Roadside Sensors for Traffic Management

Knowledge of modern state-of-the-practice traffic flow sensors provides traffic managers, researchers, and students an understanding of the operation, strengths, and limitations of current sensor technologies and enables them to make an informed decision as to which is appropriate for a particular application. Accordingly, this article describes the intrusive and nonintrusive traffic flow sensor technologies in use today, their applications and selection criteria, and typical output data. Furthermore, it provides examples of representative sensor models. The technologies discussed are mature with respect to current traffic management applications, although some may not provide the data or accuracy required for a specific application or may not perform as needed under the operational conditions anticipated at the installation site. Sensors selected for a first-time application should be field tested under conditions that will be encountered in day-to-day operation before large-scale purchases of the devices are made. As alternative traffic data and information sources, such as commercial data vendors, Wi-Fi and Bluetooth sensing of smartphone locations, and connected and automated vehicle data, become increasingly available, they are progressively finding their way into modern traffic management systems as a complement to conventional roadside sensors.


Introduction
T raffic flow sensors provide data that assist in the effective management of limited-access highways, arterials, and other city street functional classes.Such operational efficiency helps assure the safety of travelers, efficiency in moving people from one point to another (often referred to as mobility), and minimization of the environmental impacts of transportation modes.Traffic flow data are typically gathered from three types of sources: Eulerian sensors, Lagrangian sensors, and third-party vendors.
Eulerian sensors are used to monitor traffic flow at a given location and provide data that support a variety of applications, such as signalized intersection control; ramp, freeway-to-freeway, and mainline metering; wrong-way vehicle detection; queue warning; incident detection and congestion monitoring; traffic surveys; planning; and active transportation and demand management [1], [2], [3].
The most widely deployed Eulerian sensor is the inductive loop detector (ILD) that is installed in the roadway bed by a traffic management agency or contractor of its choice.ILDs provide vehicle counts; presence, passage, and lane occupancy information; and estimates of vehicle speed.Experience shows that ILDs perform well when installation and maintenance procedures that exemplify good practice are followed, such as those provided by many states, as referenced in the "ILDs" section and described in other publications [2], [4], [5].
As the variety of sensor technologies increased and matured, additional types of Eulerian sensors became available.These include improved versions of the magnetometer and magnetic sensors, which are installed in or under the roadway, and video detection systems (VDSs), microwave radar sensors, Doppler microwave sensors, passive infrared (PIR) sensors, lidar sensors, acoustic sensors, ultrasonic sensors, and sensors that employ combinations of these technologies.The latter types are installed above or to the side of the roadway, with many capable of multilane coverage.
Traffic management also relies on data from Lagrangian sensors, i.e., those that flow with the traffic [6].For example, travel times, origin-destination (OD) pair data for planning purposes, vehicle density studies, and the classification and location of congestion may be gathered from Lagrangian sensors.These sensing methods include probe vehicles [7], or floating cars, that can provide a traffic management center emissions information in addition to the usual traffic flow parameters linked to a vehicle via GPS location data [8], [9], [10] or other global navigation satellite systems' location devices, cell phone tracking through media access control address readers, automatic license plate readers, toll tag [radio-frequency (RF) identification transponder] readers, taxi fleet sources, and trucking industry transponders [11].Interested readers can find cell phone penetration rate data in O'Dea [12], a Pew Research Center fact sheet [13], and a Deloitte Touche Tohmatsu consumer survey [14].Privacy and the anonymization of cell phone user data are discussed in Herrera et al. [15], Wan et al. [16], and Daus [17].
McCracken et al. [19] discuss the opposing rationales that agencies often have for insourcing and outsourcing traffic monitoring data collection.Among the reasons for using inhouse staff are more control and confidence in data integrity.Many agencies feel that in-house staff live in and around count areas and are therefore more familiar with the historical traffic patterns of area roadways.This assists the staff to quickly identify count data that are dramatically different from historical patterns.The primary reason for outsourcing is inadequate staffing levels.Limited funding and equipment also impact in-house counts.Union rules and safety concerns for agency employees limit the ability or willingness of some agencies to send staff to the roadside for data collection.
Initiatives such as the Connected Vehicle Program in the United States [22], [23], Cooperative Intelligent Transportation Systems initiatives in Europe [24], [25], Intelligent Vehicle Innovation and Development Strategy in China [26], [27], and similar programs elsewhere are enabling vehicle-to-vehicle, vehicle-to-infrastructure, vehicle-to-pedestrian, and pedestrian-to-infrastructure communications [2].These programs utilize in-car sensors to monitor the status of vehicle systems and provide a variety of data, including braking severity, hazard warning light activation, time headways to vehicles surrounding the ego vehicle, and ego vehicle velocity, acceleration, steering wheel position, traction loss, lane departure warning, windscreen wiper activation, and air bag deployment.Public agencies and private industry are able to gather this information to gain knowledge of traffic flow rates, speeds, travel times, road weather conditions, and accident status.
Notwithstanding the importance of non-Eulerian traffic monitoring modalities, they are not discussed further as part of this review of state-of-the-practice traffic flow sensors.Nevertheless, deployment and utilization of Lagrangian sensors should be considered by traffic management agencies either as an alternative or supplement to Eulerian sensors.
The author has observed that many researchers and students refer to somewhat outdated information concerning traffic sensor operation.Accordingly, there is a need to provide an easily obtained and informed source for this vital material.This survey gives traffic management personnel, students, and researchers from the United States and worldwide access to current information concerning the operation and limitations of state-of-the-practice traffic flow sensors.In addition, it informs and aids traffic management agencies in their evaluation, selection, installation, and maintenance of the sensors.The display of test results for sensor models is minimized, especially for overhead-and side-mounted sensors, as the hardware and software for these are frequently updated, affecting the sensors' performance.Exceptions are the three-axis magnetometer and Model 702 microloop installed under the road surface, the design of which remains static.
The article is not meant to discuss experimental uses of low-cost sensors for vehicle and pedestrian detection or sensors designed for other applications, such as automated vehicle operation, vehicle diagnostics, driver assistance, or weigh in motion (WIM), to name a few.
The remainder of this article is organized as follows.The "Sensor Categorization and Selection Criteria" section describes traffic flow sensor categories and sensor selection criteria; the "ILDs" section covers ILD configurations as well as operation and utilization for speed measurement and vehicle classification; the "Magnetometer Sensors" section describes magnetometer sensors; the "Magnetic Detectors" section reviews magnetic detectors; the "VDSs" section discusses VDSs, including image processing using visible spectrum and infrared cameras and general guidelines for installing cameras at signal-controlled intersections; the "Microwave Radar Sensors" section addresses microwave radar sensors, including presence detecting microwave radar sensors, Doppler microwave sensors, radar operation, transmitted waveforms, and range and speed resolution; the "PIR Sensors" section reviews PIR sensors; the "Lidar Sensors" section concerns lidar sensors; the "Acoustic Sensors" section discusses acoustic sensors; the "Ultrasonic Sensors" section summarizes ultrasonic sensors; the "Sensor Technology Combinations" section explores sensor technology combinations; and the "Sensor Costs" section examines lifecycle sensor costs, while the "Conclusions" section summarizes the key points of the article and suggests several areas in need of further development.

Traffic Flow Sensor Categories
Conventional traffic flow sensors are often divided into two broad categories: those mounted in or under the roadway surface and those mounted above the roadway on sign bridges, traffic signal mast arms, and luminaires or to the side of the roadway on poles and other structures.The first category is also referred to as intrusive sensors because the devices infringe on the roadway pavement, while the second category is referred to as nonintrusive.
Intrusive sensors require that traffic be interrupted for installation and repair and may expose personnel to the dangers associated with working in the roadway.Installation times vary with the type of sensor.For example, ILDs typically take more time to install than magnetometers.Installation and repair of nonintrusive sensors and cameras need not interrupt traffic if the devices are mounted alongside the road.
Many traffic sensors function by detecting electromagnetic energy in some form, e.g., RF spectrum, visible spectrum, infrared spectrum, microwave spectrum, and millimeterwave spectrum.However, several of the sensors detect acoustic and ultrasonic energy.
The sensors may be either passive or active.Passive sensors only receive energy, transmitting none of their own.The received energy is a combination of energy emitted and reflected into the sensor aperture by motorized vehicles, bicycles, pedestrians, other objects of interest, the road surface, road divider structures, and extraneous sources, such as trees and other vegetation, buildings, billboard signs, and bridges.
Active sensors both transmit and receive energy.The received energy is the portion of transmitted energy scattered back into the aperture of a sensor by vehicles, the road, pedestrians, or other objects of interest and from extraneous objects.Some active sensors, such as inductive loops, detect a change in a property of the surroundings in which they are located, e.g., the inductance of the electric circuit of which the loop is a part.In others, such as microwave radars and lidars, analysis of the backscattered signal provides features that assist in classifying the vehicle or other object of interest and measuring its traffic flow parameters.Table 1 lists the sensor technologies and the categories into which they fall.

Traffic Flow Sensor Selection
The selection of a traffic sensor depends on many factors, such as those in the left column of Table 2.Not all sensors output similar data and information.For instance, sensors, in general, can provide the data in the right column of Table 2 [2], [28], [29].However, any one sensor may not provide all the data types listed.

ILDs
The ILD is the most widely used sensor found in traffic management applications.Dependent for its operation on an ac signal, this active sensor's configurations vary with  1. In-roadway-and above-roadway-mounted sensors [2], [29].
the detection objective, e.g., automobiles; scooters, motorcycles, and bicycles; long vehicles and large high-bed trucks; queue detection at limited-access highway on-and off-ramps and signalized intersections; vehicle counting; and safety and congestion applications that require speed measurements.Diamond-shaped loops also improve motorcycle detection by extending the field to the lane edges, where motorcycles sometimes drive to avoid oil spots that are more prevalent at the lane center.Bicycle detection may additionally be augmented with a loop layout resembling a figure eight [30].The most recent Traffic Detector Handbook [4] and the older edition [31] describe other loop configurations optimized for small-area detection, such as detection of a vehicle upstream of a stop line, and larger-area detection, such as presence or occupancy detection.
The Traffic Detector Handbook also discusses additional ILD design and installation topics, such as loop capacitance, the loop quality factor, loop inductance and sensitivity calculations, the number of wire loop turns required for stable system operation, effects of reinforcing steel, saw cutting   Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
operations, sealant application techniques, the splicing of the loop wire to the lead-in cable, and general loop installation guidelines [4].Moreover, it explains the operation of different types of detector electronics modules (also called detectors) that generate the ILD excitation frequency and monitor its operation.Procedures for installing ILDs in roadway pavement saw slots are found not only in the Handbook but also in ASTM's Standard Practice for the Installation of Inductive Loop Detectors [5], Klein [2], and specifications from many states that describe installation practices [32], [33], [34], [35], [36], [37], [38], [39].
The popularity of the ILD is due in part to its mature technology and low unit cost.The reliability of the wire loop has improved over the years as a result of enhanced packaging and installation practices.These include delivery of loops encased in protective materials, more thorough cleaning of debris from the sawcut, and the use of improved sealants.
The loop detector system, however, may still suffer from poor reliability.Contributing factors are poor splice connections in the pull box between the loop wire and lead-in cable; failure to twist wire pairs properly, leading to crosstalk; faulty sawcut cleaning and sealant application; and lax monitoring of the loop installation process by the responsible agency.Middleton et al. [40] and Klein et al. [4] discuss how these problems are accentuated when loops are installed in poor pavement or in areas where utilities frequently dig up the roadbed.

Operation of Inductive Loops
The inductive loop system acts as a tuned electrical circuit where the wire loops, lead-in wire, and lead-in cable are the inductive elements [4].When a vehicle passes over the loop wires or is stopped within the loop, it induces eddy currents in the wire loops, decreasing their inductance and increasing the loop system frequency.The change in frequency is sensed by the detector electronics module, which then generates a pulse output from a solid-state optically isolated device.The controller interprets the pulse as the passage or presence of a vehicle.The electronics module also contains a tuning network, oscillator, and means to ad-just the system's sensitivity.Most inductive loops operate at nominal frequencies of 15 to 100 kHz.Applications such as vehicle classification and reidentification may additionally require detector sampling rates as high as 5,000 Hz.
Loops provide accurate vehicle counts and presence indications when properly installed and maintained in good pavement.Table 3 lists the salient features of inductive loops as used for traffic management.This and similar information for other sensor technologies has been compiled over many years and documented in reports and books prepared by the author and others that are listed in the "References" section.

Speed Measurement Using Inductive Loops
Inductive loops can be utilized to measure speed in two configurations: a speed trap and as a single loop.The speed trap method depicted in Figure 1(b) is the more accurate.With this technique, vehicle speed S is calculated as where d is the distance between the leading edges of the loop pair and T T is the time difference between the pulses produced by the electronics module when a vehicle is detected by the leading edge of the first loop and vehicle detection by the leading edge of the second loop.
A study by Woods et al. [41] indicated that the optimal distance between loops is 9 m (30 ft).Closer spacing, such as 6 m (20 ft), can cause crosstalk between two identical detectors operating on the same frequency and sensitivity, creating a conflict with the requirement for accurate speed measurement.
Their findings emphasize three criteria for accurate speed measurement using a speed trap configuration of loops: first, the loops must be identical, using multiconductor cable; second, identical detector electronics modules must be used and operate on the same frequency and sensitivity settings; and third, the loop pairs must be calibrated on a regular schedule.
Single loops can be utilized to measure vehicle speed when a less accurate estimate is acceptable.Coifman [42]   and Lu et al. [43] analyzed the factors that affect the estimate of the average length of a vehicle passing over a loop, a parameter required by the single-loop speed estimation technique.Vehicle length values and accuracies can vary by lane and vehicle mix (trucks are more frequently found in the rightmost two lanes of a roadway), road location (urban or rural), type of road (arterial or limited access), season, time of day, weather, and occurrence of special events.
Other information required to implement single-loop speed measurement includes measured values of vehicle count and occupancy and the effective loop length.The effective loop length may differ from the physical dimensions of a loop, as detection may occur as soon as a vehicle enters the electromagnetic field of the loop, which lengthens the effective detection area.
Average vehicle speed S r from the single-loop method is given by where S r is the speed in miles per hour, 0.6818 is a constant that converts feet per second into miles per hour, VC is the vehicle count during the measurement period, LL is the effective loop length in feet, VL is the vehicle length in feet, and O is lane occupancy during the measurement period measured in seconds.
The state of the practice in ILD technology provides single-loop speed estimates that differ from the true value by as much as 30%.To obtain even these relatively crude measurements, the vehicle count, vehicle length, and occupancy must be known to within an error no greater than %. 10 !Of these, vehicle length is the most difficult to estimate accurately for the reasons mentioned previously.The allowable error in vehicle length is discussed further by Klein [2], [28].

Vehicle Classification Using Inductive Loops
Minge et al. [44] and Yu et al. [45] describe traditional ILD vehicle classification methods that rely on axle and length measurements.Nowadays, microprocessors enable ILD electronics modules to directly classify vehicles by collecting and analyzing signatures of a vehicle's undercarriage.
For example, the I-Loop Duo card in Figure 2 is capable of classifying the traffic stream into the first 13 U.S. Federal Highway Administration (FHWA) classes [46].The 13 classes are based on whether a vehicle carries passengers or commodities.Nonpassenger vehicles are further subdivided by the number of axles and number of units, including both power and trailer units.In addition to traffic flow rate, vehicle presence, and classification data, the card's sample rate of 100-5,000 Hz supports measurement of travel time, vehicle speed, and OD pairing through  downstream reidentification of the unique signature generated by a vehicle's undercarriage [49].
In a particular application of the I-Loop card to vehicle classification, the card's high-resolution signatures of a vehicle's metal undercarriage are integrated with WIM system classification attributes.Hernandez et al. [47] demonstrated that this combination of features allows categorization of trucks into 14 classes based on axle count, axle spacing, and gross vehicle weight.
Another study of the I-Loop card's classification accuracy by Liao [48] in the Minneapolis-St.Paul, MN, USA, Twin Cities region indicated that damaged loops, broken loop sealant, crosstalk, and not twisting the lead-in cable properly could influence the ability of the card to accurately classify vehicles.

Magnetometer Sensors
Magnetic sensors are passive devices that indicate the presence of a metallic object by detecting a perturbation (known as a magnetic anomaly) in Earth's magnetic field created by the object.Two types of magnetic field sensors are found in traffic flow parameter measurement: magnetometers and magnetic detectors.They differ operationally in that the first type can detect stopped vehicles, while the second category cannot.
Three-axis fluxgate magnetometers detect changes in the x, y, and z components of Earth's magnetic field produced by a ferrous metal vehicle.They contain a primary winding and secondary "sense" windings on a bobbin surrounding a high-permeability soft magnetic material core.
The magnetometer's secondary windings generate an output voltage in response to the magnetic field anomaly created by a vehicle.The sensor declares a vehicle present when the voltage exceeds a predetermined threshold.In the presence mode of operation, the detection output is maintained until the vehicle leaves the detection zone.
Magnetometer applications include traffic-adaptive signal control, ramp metering, limited-access highway data collection, and parking occupancy detection as well as on bridge decks where ILDs may be affected by the steel support structure or simply cannot be installed because of other constraints.Arrays of three-axis fluxgate magnetometers can gather vehicle signatures in support of vehicle classification and travel time notification.Magnetometer sensors, such as those in Figure 3, provide vehicle flow data, such as presence, passage, count, and lane occupancy.Since magnetometers are wirelessly deployed, they reduce installation time and cost as compared to ILDs.
When installed in a speed trap configuration, magnetometers may be used to measure vehicle speed.Some applications may require more than one sensor across a lane to guarantee 100% vehicle detection, such as at a signal stop line to detect motorcycles at a lane edge or to determine queue length, as Sanchez et al. [50] showed for a curved section of road where vehicles were partially in one lane and partially in another.
• Sensing area is approximately that of a 2-m (6-ft)-diameter round loop.
• Flush-mounted sensor: Installed by coring a 10-cm (4-in)-diameter × 7-cm (2.75-in)-deep hole, inserting the sensor into the hole, aligning it with the direction of traffic flow, and sealing the hole with fast-drying epoxy.• 10-year battery life based on 300 million detections.
• Base radio at an intersection receives sensor data to 46 m (150 ft).Repeater modules extend range up to 610 m (2,000 ft).
• Up to 300 m (1000 ft) data transmission range.
• Spread spectrum and time-division multiple-access RF communication and control.
• Installed in a 7.6-cm (3-in)-diameter × ~7.6-cm-deep hole.Yang and Zuo [51] demonstrated that inclement weather, such as rain, fog, mist, haze, or snow, may increase stuckon call errors, while the frequent passing of heavy vehicles at large intersections could interrupt wireless communication between the magnetometers and roadside access point that receives the magnetometer data, thus increasing false and stuck-on call errors.The interruption in communication can be resolved by installing additional repeaters.
There appears to be interest in installing magnetometers at the side of a roadway for vehicle counting, speed, and classification [52], [53], [54].Such configurations eliminate the need for interrupting traffic flow, may reduce the safety measures needed for installation personal, and may make removal of debris from the coring operation simpler.However, this sensor configuration is effective only at measuring traffic parameters in the lane immediately adjacent to the installation site.
Middleton et al. [55] compared magnetometer vehicle counts with ILD counts at two sites.Their findings indicated agreement within 1% to 2% at one location and 0.1% at the other.Estimates of sensor reliability, initial costs, and user-friendliness are also noted in this report.

Magnetic Detectors
The second type of magnetic field sensor, the magnetic detector, is more properly referred to as an induction coil magnetometer or search coil magnetometer [4].It detects a vehicle signature by measuring the deformation in the magnetic flux lines induced by the change in Earth's magnetic field from a moving ferrous metal vehicle.These devices contain a single coil winding on a permeable magnetic material rod core.Similar to the fluxgate magnetometer, magnetic detectors generate a voltage when a ferromagnetic object perturbs Earth's magnetic field.However, most magnetic detectors do not detect stopped vehicles since they require a vehicle to be moving or otherwise changing its signature characteristics with respect to time.
Nonetheless, multiple units of some magnetic detectors can be installed and utilized with manufacturer-supplied signal processing software to generate vehicle presence data.
Figure 4 illustrates the Model 231 magnetic detector system components (the probe and either a Model 201 or 232 amplifier connected to the probe's output signal in a controller cabinet) and the probe's installation by insertion into a trench under the roadway.
The Model 701 microloop sensor, pictured in Figure 5(a), is inserted vertically in 2.5-cm (1-in) holes and placed 46 to 61 cm (18 to 24 in) below the roadway surface.Up to four 701 sensors can be connected in series.The Model 702, depicted in Figure 5(b), is an example of a magnetic detector that can be utilized with manufacturer-supplied signal processing software to generate vehicle presence data.It is inserted into a 7.6-cm (3-in) nonferrous schedule 80 conduit.The conduit is then buried 53.3 !7.6 cm (21 ! 3 in) below the road surface using horizontal directional drilling or open trenching techniques.
Middleton and Parker [56] provide cost and vehicle count as well as speed accuracy data for the Model 701 and 702 sensors, while Minge et al. [57], [58] offer similar information for the Model 702.Grone [59] investigated the accuracy of the Model 702 for presence detection and vehicle classification based on length.The results of these studies are conveyed in Table 4.
Table 5 summarizes the prominent attributes of magnetic sensors as used for traffic management.Now that we have examined sensors that are mounted in or under the roadway bed, let us explore sensors that are mounted above or to the side of the roadway.

VDSs
Video cameras were introduced to traffic management for roadway surveillance based on their ability to transmit closed-circuit television imagery to a human operator for interpretation.Later, they were incorporated as the sensing element in VDSs, such as those in Figure 6.These passive sensors produce imagery from energy emitted and reflected from objects in their field of view into their aperture.Traffic management applications utilize image processing to analyze a scene of interest and extract information for traffic surveillance; signal control; incident detection on streets, on freeways, and in tunnels; wrong-way driver detection; and ramp and freeway metering, to name a few.Signalized intersection control is the most prevalent application, according to several VDS manufacturers.
A VDS typically consists of one or more cameras, a microprocessor-based computer for digitizing and analyzing the imagery, and software for interpreting the images and converting them into traffic flow data.Such a system can replace several in-ground inductive loops, provide detection of vehicles across several lanes, and perhaps lower maintenance costs.Some systems process data from more than one camera and thus further expand the area over which data are collected.
Older models provide limited vehicle classification by length, while newer models with edge processing utilize artificial intelligence (AI) deep learning algorithms for image classification.VDSs report vehicle presence, flow rate, lane occupancy, and speed for each class and lane.The ability to position the camera with sufficient mounting height at an optimized mounting location for a given application impacts the data measurement accuracy.For these reasons, most manufacturers will not provide a confidence interval for their measurement accuracy specifications [60].Klein [2, Ch. 9] elaborates on the advantage of pairing a confidence interval with an accuracy specification.

VDS Image Processing
Intersection-based applications of a VDS may utilize differently shaped detection zones, e.g., down lane, stop line, and speed detection zones, to differentiate between vehicles and bicycles and collect various types of data [61], [62].Other systems track vehicles through the entire field of view of the camera, using Kalman filtering algorithms to update vehicle position and velocity estimates.By processing trajectory data obtained from the time trace of the position estimates, local traffic parameters (e.g., vehicle flow rate, lane change frequency, and turning movements) can be computed.
Changes in traffic scene imagery between successive frames also provide information that assists in computing traffic flow data.These data, together with other vehicle identifiers (e.g., time stamp, vehicle type, color, shape, position, and speed), can be communicated to the traffic management center to implement traffic control measures.Tracking vehicle subfeatures, such as edges, corners, and 2D patterns, rather than entire vehicles has been proposed to make VDSs robust to partial occlusion of vehicles in congested traffic.
VDSs function adequately under most operational conditions and show performance improvement over time.Some VDSs contain software that limits the effects of wind-induced camera movement artifacts, glare, and poor lighting conditions.
The first types of image processing algorithms analyzed black-and-white imagery to examine the variation of gray levels in groups of pixels (picture elements) contained in the video frames.These algorithms remove gray-level variations in the image background related to fixed objects (e.g., guardrails, trees, shrubs, and buildings), weather conditions, shadows,   and daytime or nighttime artifacts and retain objects identified as automobiles, trucks or buses, motorcycles, bicycles, and pedestrians.Chromatic information from color imagery enhances vehicle discrimination in inclement weather and when camera mounting conditions are not ideal, assists in differentiating vehicles from shadows, and helps in identifying features of individual and groups of vehicles.VDS manufacturers may obtain vehicle speed and classification by measuring and calibrating the field of view by using traffic cones to mark appropriate distances.Alternatively, they may employ an algorithm that simply determines small, medium, and large vehicles [60].
Deep learning artificial neural networks are increasingly employed to identify vehicles and pedestrians in imagery obtained from roadway-mounted cameras.For example, Razi et al. [63] describe newer approaches to image processing and identification of vehicles and pedestrians using deep learning artificial neural networks that utilize  multilayer perceptrons, region-based convolutional neural networks, and recurrent neural networks.
Patrikar and Parate [64] discuss the properties of edge detection computing algorithms that detect and classify vehicles and other objects.Edge AI brings computational power closer to the data source, allowing on-device data processing and enabling real-time situation-aware decision making.This architecture limits the amount of data transferred to the cloud or traffic management center for analysis and thus supports reduced latency, improved bandwidth efficiency, enhanced data privacy, and increased reliability in scenarios with limited or intermittent connectivity.However, the types of algorithms that can be implemented may be limited by the throughput and memory capabilities of the microprocessor, and there may be fewer camera options to choose from.
Internet protocol cameras that incorporate edge processing are finding their way into an increasing number of VDS models.These cameras eliminate the need for cabinet-based processors and can result in lower-cost detection systems.
Classical thermal imaging sensors utilized in military applications defined classification as the set to which a vehicle belonged (e.g., automobile, pickup truck, motorcycle, 18-wheeler, or bus), while identification was reserved to denote a vehicle's description up to the limit of the sensor's ability to differentiate one vehicle from another.It included attributes such as the vehicle's manufacturer and model and perhaps color (e.g., Toyota Corolla, Chevrolet Equinox electric vehicle, or Mercedes-Benz GLE sport utility vehicle).Generally, a higher resolution and signal-to-noise ratio are required to perform identification as compared to classification [65], [66], [67], [68], [69].
Modern AI vehicle recognition deep learning algorithms use another lexicon.Here, the term recognition is used to describe a collection of related computer vision tasks whose purpose is to characterize objects in digital imagery [70], [71], [72], [73], [74], [75].The tasks encompassed by the term may vary among researchers.The tasks often include image classification, detection, semantic segmentation, instance segmentation, tracking, and restoration, defined as follows: ■ Image classification assigns a predefined label to the input pixels and distinguishes object categories.
■ Object detection predicts the location and extent of each object by utilizing bounding boxes.
■ Semantic segmentation predicts and assigns a specific category label to each object but does not distinguish among multiple objects of the same category.
■ Instance segmentation tags each specific object in a given object category down to the pixel level.
■ Tracking locates and follows moving objects to aid in decision making.
■ Image restoration converts low-resolution pixels into high-resolution ones.

Infrared VDS
As infrared cameras become cost competitive, they are being utilized by traffic management agencies to detect the heat signatures of vehicles, bicycles, and pedestrians, as those do not produce the glare that often accompanies visible spectrum imagery.In addition, longer-wavelength infrared cameras operate in darkness or poor lighting conditions and through smoke and light fog.They can also detect vehicles that may be difficult to distinguish in shadows.Applications include wrong-way vehicle detection by the Arizona Department of Transportation, pedestrian detection at intersections in Qatar, and incident detection in Istanbul, Turkey [76].
Infrared image features, whether for automobile, bicycle, or pedestrian detection, are created by reflected and emitted energy captured by a camera operating in an infrared wavelength band.There are three commonly cited infrared bands: near-infrared, from 0.87 to 1.5 µm; midinfrared, from 3 to 5 µm; and long-wavelength thermal energy that extends from 8 to ≥12 µm.By contrast, features in visible spectrum (0.4 to 0.7 µm) images are formed by reflected sunlight or light captured from headlights and taillights that enters the camera lens.
Images in the near-infrared band are generated predominantly by reflected energy and appear similar to visible wavelength images to the human eye.Midinfrared images begin to take on emissive characteristics, where features are proportional to the emissivity of the radiating surface and its absolute temperature.Long-wavelength infrared images are predominantly formed by energy emitted from the objects in the field of view of the camera.These images have a small reflected light component and appear different than visible wavelength images to the human eye.MacCarley et al. [77] and Grossman et al. [78] point out that long-wavelength, or thermal infrared, cameras are not subject to sun glint and the effects of inadequate lighting as are video systems operating in the visible spectrum.
Figure 7 displays an infrared camera detection system designed for intersection signal control, wrong-way direction detection, and bicycle and pedestrian detection.The sensor initiates the opening and closing of a pair of isolated contacts in the controller cabinet in response to a vehicle, bicycle, or pedestrian detection.The contact closures provide information concerning the number of vehicles passing the sensor per hour or the presence of a motorized vehicle, bicycle, or pedestrian in the detection area of the sensor, as illustrated on the right side of the figure.

General Guidelines for Installing VDS Cameras
Four general guidelines for installing single-camera VDSs at signalized intersections are as follows [2], [60], [61], [79], [80], [81]: 1) Maximize the camera height (25 to 30 ft is a common requirement, although some systems recommend heights Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply. as high as 40 ft) to minimize vehicle occlusion and headlight reflection artifacts and thus maximize the accuracy of vehicle count, speed, and presence measurements.2) Select the camera location to avoid or minimize occlusion of vehicles in the monitored lanes.3) Do not include the horizon in the camera view.4) Adjust the sun shield to minimize glint during sunrise and sunset with east-west-facing cameras.
Vehicle occlusion can manifest in several ways.It may be caused by a lack of sufficiently large gaps between vehicles in lanes closest to a side-looking camera that obscure vehicles in the more distant lanes.It may also be produced by the blocking of distant vehicles by a long or tall vehicle in nearby lanes.
Other VDSs, such as Gridsmart, Miovision, and Cyclops, utilize a dome camera and consequently have another set of installation requirements.Dome cameras installed for intersection monitoring often require a second camera for servicing larger intersections and another sensor for advance vehicle detection.
Modern high-definition cameras (720 or 1,080 pixels) have larger fields of view than older models.Therefore, they can eliminate the need for additional cameras at the same approach when monitoring left-turn pockets or detecting receding vehicles to obtain departing traffic counts [60].
When operating a VDS at night, adequate street lighting should be provided to ensure reliable vehicle detection by the VDS [82].Most VDSs issue a recall to the controller if a vehicle is not detected within some specified time period to prevent the vehicle from being trapped at the intersection.
Table 6 summarizes the prominent characteristics of a VDS as used for traffic management.Several limitations noted in past tests [55], [57], [58], [83], [84], [85], [86], [87], [88] may not be present or as pronounced in newer systems due to improvements in the signal processing algorithms and hardware.However, the weaknesses are worth noting  so that potential users of these devices can make an informed purchase and determine whether field tests and installation enhancements, such as lighting and removal of obstacles in the camera's field of view, are required.

Microwave Radar Sensors
Two types of microwave sensors are utilized in traffic management applications-presence detecting and Dopplertypical models of which appear in Figures 8 and 9, respectively.These sensors are active devices that transmit energy and receive the portion that is scattered back into their aperture, in this case, an antenna.Presence detecting radar models detect stopped vehicles and measure vehicle speed [89], while the Doppler models usually require the speed to be greater than some minimum value for vehicle detection.Presence detecting radars find application in signalized intersection control, especially as advance sensors, wrong-way vehicle detectors, and highway incident detectors.Doppler microwave sensors measure vehicle speeds on many limited-access highways.On arterials and other classes of city streets, they monitor vehicle speeds, warn drivers of excessive speed, and alert work zone personnel to oncoming vehicles.
Estimates of the accuracy, sensor reliability, initial costs, and user-friendliness of earlier versions of presence detecting microwave radar sensors are found in Middleton  and Parker [56] [84] and Middleton et al. [55].Yu and Prevedouros [88] also determine the accuracy of a presence detecting microwave radar, while Mohammed [90] assesses accuracies of a presence detecting microwave radar and a Doppler microwave sensor.
Microwave radars and Doppler sensors have shown improved resolution and accuracies with time.Furthermore, many provide per-vehicle data that support new FHWA annual average daily traffic reporting standards [91] and user interfaces that simplify setup and calibration.

Presence Detecting Microwave Radar Sensors
Most present-day presence detecting microwave radar sensors operate in the 24-GHz band, although some function in the 10-GHz band.They differentiate vehicles in multiple lanes from a side-looking or forward-looking configuration, depending on the model.The models in Figure 8 are meant to be representative of products and capabilities that are currently offered.Not all available versions are included, for example, those that detect pedestrians waiting to cross an intersection.Manufacturers or distributors of these and other devices should be contacted by interested purchasers to learn of the latest sensor models, features, installation procedures, and costs.

Doppler Microwave Sensors
Doppler microwave sensors also operate in the 24-GHz band but generally do not detect stopped vehicles because they rely on the Doppler principle to sense a vehicle; i.e., the vehicle must be moving at a speed greater than some minimum established by the manufacturer.For example, Doppler microwave sensors may not be suitable for measuring vehicle speed under congested traffic conditions, as they typically do not detect vehicles traveling at less than the specified minimum, e.g., 3-11 km/h (2-7 mi/h).
With some ingenuity, these sensors can be made to report the positions of slow-moving and stopped vehicles using algorithms that track vehicle speed trends.For instance, if a Doppler sensor is tracking a group of vehicles whose speed is decreasing, at some point, the sensor will not report the vehicles' speeds nor their location.In this case, an algorithm can be written to infer that the vehicles in question have stopped moving or slowed below the minimum detectable speed but are still on the roadway.The models in Figure 9 measure speed and other characteristics of vehicles in support of various applications.

Radar Operation
Radar detection is a stochastic process; i.e., the radar sensor may or may not produce a detection event when a vehicle is present in the sensor's field of view.A radar sensor may also produce a detection event when an object other than a vehicle is in the field of view of the sensor, e.g., clutter or some other object not of interest, such as metallic structures on a bridge or fence.The latter detections are called false alarms and may be reduced by requiring the signal corresponding to valid vehicle detections to exceed a threshold voltage and contain a number of characteristics common to those of the vehicles of interest.The probability of a valid detection is specified through a detection probability and the probability of a false detection event through a false alarm probability.These parameters are a function of the radar design and data processing.
Forward-looking radars with large antenna beamwidths acquire data representative of the composite traffic flow in one direction over multiple lanes.Forward-looking radars with narrow antenna beamwidths generally monitor a single lane or several individual lanes of traffic flowing in one direction, depending on the model.Side-looking multiple-detection zone radars project their detection area (i.e., footprint) perpendicular to the traffic flow direction and report traffic data from several individual lanes.

Transmitted Waveforms
where f is the transmitted frequency, fD is the Doppler frequency, c is the speed of light, and i is the angle between the direction of propagation of the sensor energy and the direction of travel of the vehicle.The frequency of the received signal is , f fD !where !denotes whether the vehicle is moving toward or away from the sensor.Accordingly, the received signal frequency increases when a vehicle moves toward the sensor and decreases when a vehicle moves away from the sensor.Vehicle passage or count is denoted by the presence of the frequency shift.Vehicle presence cannot be measured with the constant-frequency waveform, as only moving vehicles are detected by most sensors of this type.
Figure 10(b) presents the waveform transmitted by many presence detecting microwave radars.This frequencymodulated continuous-wave (FMCW) signal is characterized by a transmitted frequency that is constantly changing with respect to time.It is this feature that enables the radar to measure range and detect vehicle presence, i.e., stopped vehicles.The range R to the vehicle is proportional to the difference in the frequency f T of the transmitter at the time t1 when the signal is transmitted and the time t2 when a portion of the transmitted energy is received back at the sensor or, equivalently, the time difference -This relation appears as where c is the speed of light.
The equation for range may also be written in terms of other FMCW radar design parameters as where f T is the instantaneous difference in the frequency of the transmitter the times the signal is transmitted ( ) t1 and received ( ), t2 F T is the RF modulation bandwidth (a design parameter), and fm is the RF modulation frequency (another design parameter).

Range and Speed Resolution
Range resolution , R T the minimum distance resolved by an FMCW radar, is given by Therefore, if the radar sensor operates in the 24-GHz band with 75 MHz of RF bandwidth, it has a range resolution of 2 m (6.6 ft), while one with 245 MHz of RF bandwidth produces a range resolution of 0.61 m (2.0 ft).
Speed or Doppler resolution fD T is where fm is defined above and Tm is the reciprocal of .fm The range bins in a forward-looking sensor operate similarly to the speed trap configuration of inductive loops for measuring vehicle speed.When a vehicle enters the first range bin, a pulse is created by the signal processing electronics.A similar pulse is formed when the vehicle enters the second range bin.The vehicle speed is found by dividing the distance between the range bins (a known design quantity) by the time difference between the start of the first and second pulses [2], [28], [29].
Dual-antenna radar models installed in a side-looking configuration measure vehicle speed in an analogous manner using the travel time between the two receive antenna beams that are projected perpendicular to the direction of travel.These devices do not require calibration for speed measurements.
Side-looking radars with a single antenna measure speed using the same principle as a single inductive loop and thus require a form of calibration.Here, an assumed vehicle length is needed along with a vehicle occupancy measurement.
Table 7 describes the characteristics of presence detecting microwave radar sensors as used for traffic management.Table 8 contains similar information for microwave Doppler sensors.

PIR Sensors
Figure 11 provides an example of a PIR sensor.These devices detect energy from two sources: 1) energy emitted from vehicles, road surfaces, and other objects in their field of view and 2) energy radiated by the cosmos, galaxy,  [68], [92], [93], [94] Lower-frequency models penetrate foliage With side mounting, possibility of missed detections if tall vehicles occlude more distant lanes when congestion is heavy; therefore, some models not recommended for stop line detection (may not be a significant effect in other applications) Vehicle undercounting may increase in heavy congestion Setback mounting distance must be accommodated Stochastic property of radar detection may cause device to miss detecting a vehicle (using multiple-detection zones and device design features, such as a high FMCW repetition frequency, may ameliorate this issue) Table 7.The presence detecting microwave radar typical output data, installation location, advantages, and limitations.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
and atmosphere that is reflected by vehicles, road surfaces, or other objects into the sensor aperture.They transmit no energy of their own.An optical system focuses the detected energy onto an infrared-sensitive material, called the detector, which is mounted at the focal plane of the optics.The detector converts the emitted and reflected energy into electrical signals, which are analyzed in real time to determine the presence of a vehicle and vehicle flow rates, speeds, and classes.Pyroelectric materials, such as lithium tantalate, are common detectors found in PIR sensors that acquire traffic data.These sensors are utilized for signal control, pedestrian detection, axle counting, and transmission of traffic information to motorists.
Multichannel (i.e., more than one type of sensor technology) and multizone (i.e., more than one detection region) PIR sensors measure speed, vehicle length, flow rate, and lane occupancy.Multizone models are designed with dynamic and static thermal energy detection zones that provide the functionality of two inductive loops.Time delays among the signals from three dynamic zones are required to measure speed.The vehicle presence time from a fourth zone is used to calculate the lane occupancy of stationary and moving vehicles.
Cost considerations make the infrared band a good choice for vehicle sensors that incorporate a limited number of pixels.Most models operate in the long-wavelength infrared band from 8 to 14 µm and thus minimize the effects of sun glint and changing light intensity from cloud movement.The sensor's electronics distinguish between energy detected from the road surface when no vehicle is present and energy detected when a vehicle is present and thus are able to report a vehicle presence event.

Planck's Radiation Law
Energy is emitted at all frequencies by objects not at absolute zero (-273.15°C).The emitted energy is governed by the Planck radiation law and is a function of the received frequency and surface temperature of the object of interest.
With perfect emitters, or blackbodies, an object's surface physical temperature T is equal to the brightness temperature TB measured by a PIR vehicle sensor.How-ever, the surfaces of real objects do not normally radiate as blackbodies (i.e., they are not 100% efficient in emitting the energy predicted by the Planck radiation law).To account for this nonideal emission, a multiplicative emissivity factor is added to more accurately predict the amount of energy radiated by the object, now referred to as a gray body [93], [95].Emissivity f is equal to the ratio of TB to T, where .0 1 # # f PIR sensors detect not only emissions from a vehicle or road surface in their field of view but also cosmic, galactic, and atmospheric radiation.These latter effects are often combined into one term, referred to as "sky radiation" or "sky temperature." Overhead-mounted PIR sensors detect radiation from two pathways.The first is emission generated by the atmospheric region between the downward-looking sensor and road surface.The second arises through detection of the reflected cosmic-, galactic-, and atmospheric-radiated energy from surfaces whose emissivity is not unity, e.g., metallic vehicles and, to a lesser extent, the road surface.Perfectly reflective objects have an emissivity of zero.The first contribution is small compared to the reflected energy component, due to the relatively short ranges at which PIR sensors operate in most traffic management applications.Additional information concerning the utilization of radiative transfer theory for vehicle detection is available in Klein [2], [28].

PIR Sensor Summary
Table 9 lists the output data, installation location, advantages, and performance limitations of PIR sensors.Several disadvantages of PIR sensors are sometimes cited.Glint from sunlight may cause unwanted and confusing signals in the shorter PIR sensor wavelength bands.Atmospheric particulates and inclement weather can scatter and absorb energy that would otherwise reach the focal plane of the sensor.The scattering and absorption effects are sensitive to water concentrations in fog, haze, rain, and snow and the presence of other obscurants, such as smoke and dust [67], [95], [96].Since traffic management applications encounter relatively short operating ranges, these concerns may not be significant.However, some performance degradation (e.g., undercounting) in heavy rain and snow has been reported.

Lidar Sensors
Lidar sensors are active devices that illuminate a roadway surface with energy transmitted by laser diodes operating in the near-infrared region of the electromagnetic spectrum at wavelengths of 0.905 to 0.94 µm.They detect the portion of the transmitted energy that is reflected or scattered by vehicles back toward the sensor.The sensors are mounted overhead or to the side of the roadway to view approaching or departing traffic, as exemplified by the models in Figure 12.Side-looking configurations may offer coverage of additional lanes and axle counting.Lidars provide vehicle presence at traffic signals, flow rate, speed, length assessment, queue measurement, and classification, with some models providing up to 30 defined vehicle classes [97] in support of toll collection and other applications.Lidars project their detection zones across a lane in various ways.In the first technique, the transmitting optics divide the pulsed laser diode output into two beams separated by several degrees.A rotating mirror scans the beams across a lane.Speed measurement occurs by recording the times at which a vehicle enters the detection area of each beam.Since the beams are a known distance apart, the speed is given by the ratio of the distance to the difference corresponding to the vehicle's arrival at each beam.The LaserScan sensor in Figure 12(a) is typical of this design [98], [99], [100].
Sensors manufactured by SICK also scan the laser beam with a rotating mirror.However, the speed calculation is unlike that of the LaserScan and depends on whether the system contains one or two lidars [97].
SICK one-lidar systems shown in Figure 12(b) and (c) calculate speed from a vehicle's width measurement.The width is found by detecting the diagonally opposite corners, e.g., left-front and right-rear corners, of the vehicle as it passes through the field of view of the lidar.Speed is computed as the width measurement divided by the time difference between the detection of the vehicle's corners.
In two-lidar SICK systems, the laser beam is projected along the direction of travel to detect the top, front, and rear of a vehicle.The vehicle length and speed are computed from these data and the time difference in the measurements corresponding to the front and rear of the vehicle.
The second method of laser beam projection across monitored lanes is found in lidars that utilize several laser diodes and detection elements to transmit and receive multiple side-by-side beams and thus do not require a scanning mirror to illuminate an entire lane.
Still other lidars are being developed to provide highresolution imagery of intersections.These systems use several lidars to scan an entire intersection to detect and classify motorized vehicles, bicycles, and pedestrians and notify traffic management personnel of turning movements and other pertinent information [101].
Table 10 summarizes the output data, installation location, advantages, and performance limitations of lidar sensors.Snowy and rainy days and other challenging weather conditions affect the accuracy of lidar vehicle detection [102].A rule of thumb for determining when a sensor operating in the near-infrared wavelength band, such as lidar sensors, may experience difficulty detecting a vehicle    under adverse conditions is to note whether a human observer can see the vehicle under the same circumstances.If the observer can see the vehicle from the operating range of the sensor, there is a high probability that the lidar will detect the vehicle as well.

Acoustic Sensors
Acoustic sensors utilized in traffic management are passive devices that transmit no energy of their own.Passive acoustic sensors measure vehicle passage, presence, and speed by detecting acoustic energy, or audible sounds, produced by vehicular traffic from a variety of sources, such as engines and the interaction of a vehicle's tires with the road.When a vehicle passes through the detection zone, the signal processing algorithm responds to an increase in sound energy, and a vehicle presence signal is generated.When the vehicle leaves the detection zone, the sound energy level drops below the detection threshold, and the vehicle presence signal is terminated.Figure 13 displays an acoustic sensor that incorporates a microphone array and signal processing algorithms that impart spatial directivity to detected sounds and associates them with vehicles traveling in different lanes.Signals emanating from locations outside the detection zone are attenuated and ignored [103].
Middleton and Parker [56], [84] provide cost, vehicle count, and speed accuracy data for the SmarTek SAS-1 acoustic sensor from evaluations conducted in 2000 and 2002, respectively.Since the sensor has undergone improvements from continued testing and deployments, the values reported may not reflect present-day accuracies.Table 11 itemizes the output data, installation location, advantages, and performance limitations of passive acoustic sensors.

Ultrasonic Sensors
Ultrasonic sensors are active sensors that transmit pressure waves of sound energy at a frequency between 25 and 50 kHz, which is above the human audible range.The most accurate data are obtained when the sensors are mounted over the center of the monitored lane.An alternate mounting location at the lane edge (especially if the monitored lane is the rightmost lane) is sometimes used.The sensors can also be mounted in a horizontal position when used as vehicle detection triggers, for example, to prevent a barrier gate in a parking structure from closing on top of a vehicle.
The vast majority of ultrasonic sensors transmit pulse waveforms that provide vehicle count, presence, and occupancy information.The sensor measures distances to the road surface and vehicle surface by detecting the portion of the transmitted energy that is reflected toward the sensor from an area defined by the transmitter's field of view.When a distance other than that to the road surface is measured, the sensor interprets the measurement as the presence of a vehicle.The received ultrasonic energy is converted into electrical energy, which is analyzed by signal processing electronics either collocated with the transducer or placed in a roadside controller cabinet.
The transmitted pulsewidths Tp are typically 0.02 to 2.5 ms, with repetition periods T0 (the time between bursts of pulses) typically 33 to 170 ms.The sensor measures the time it takes for the pulse to arrive at a vehicle and return to the transmitter.The receiver is gated on and off with a user-adjustable interval that differentiates between pulses reflected from the road surface and those reflected from vehicles.The detection gates of various models adjust to support vehicle Poor foliage penetration Mean time between failures is reduced in models that use rotating mirrors at the larger scan frequencies.However, all models provide adequate performance for their intended applications.
Table 10.The lidar sensor typical output data, installation location, advantages, and limitations.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
detection at distances greater than .0 5 + to 0.9 m above the road surface.A hold time T h (composite values from manufacturers range from 115 ms to 10 s) is built into the sensors to enhance presence detection.
Constant-frequency ultrasonic sensors that measure speed using the Doppler principle are manufactured in Japan to interface with its highway infrastructure.However, these are more expensive than pulsed models.They mount overhead, facing approaching traffic at a 45° incidence angle.The sensor has two transducers, one for transmitting and one for receiving a signal.Vehicle passage is detected by a shift in the frequency of the received signal.Vehicle speed can be calculated from the pulsewidth of an internal signal generated by the sensor's electronics that is proportional to the speed of the detected vehicle [104].
Table 12 lists the output data, installation location, advantages, and performance limitations of ultrasonic sensors.Temperature change and extreme air turbulence may affect their performance.Temperature compensation is built into some models.Large pulse repetition periods may degrade occupancy measurement on limited-access highways when vehicles are traveling at moderate to high speed, as an insufficient number of pulses is transmitted and reflected from a vehicle while it is in the sensor's detection zone.

Sensor Technology Combinations
Several manufacturers combine two or more technologies in a single sensor unit for specialized applications or enhanced operational performance.Figure 14 displays infrared-ultrasound and infrared-Doppler-ultrasound sensors along with their principal characteristics [105], [106].Furthermore, several of the VDS models in Figure 6 combine video and radar technologies [107].

Sensor Costs
The higher cost of above-roadwaymounted sensors is often offset by the costs associated with installing and maintaining multiple lower-cost sensors, such as inductive loops.Cost estimates for magnetometers, magnetic detectors, and presence detecting microwave sensors appear in Middleton et al. [40], [55], Middleton and Parker [56], [84], and Minge et al. [57], [58].Several of these contain purchase and installation costs.However, these values were obtained over a decade ago and, furthermore, pertain to locations where labor and material costs may differ materially from those in another area.Since the costs may not be representative of presentday values, they are not cited here.
Table 13 provides a range of costs to account for differences among models and optional hardware in each technology group, cables, discounts that may be available for quantity purchases, and the number of lanes monitored.
The cost for microwave radars is indicated per approach.This is typical of vendor cost estimates for most models intended for traffic signal control.One model does feature the ability to monitor two approaches with a single unit.The limited-access highway incident detection application usually requires only one radar per direction of travel.
Comprehensive evaluations of sensor costs must account not only for the sensor purchase but also software and associated updates, training, software initialization and detection-zone setup, required cables, installation and first-time use testing [2], [108], [109], [110], data transmission, maintenance, and other costs associated with the implementation of lane closures, safety measures, environmental hazard mitigation, preparation of the road surface or subsurface for inductive loops or other surface or subsurface sensors, and any mounting structures and power sources that must be added to the infrastructure.
Maintenance and repair estimates may be available from manufacturers and other agencies and localities   that have deployed similar sensors.Some above-roadwaymounted sensors are designed with a mean time between failures of 35,000 to 90,000 h [97], [111], [112].

Conclusions
Knowledge of the theory of operation of modern traffic flow sensors offers traffic management personnel and researchers the ability to appreciate the installation requirements, strengths, and limitations of each sensor technology option and thus make an informed selection for a particular application.The inductive loop, magnetic, and above-roadway-mounted sensor models appearing in this article are typical of those that support current traffic management applications.Since new sensors are constantly reaching the market, existing capabilities may be superseded by those of newer models or by the introduction of new products by other manufacturers.Current sensor options and installation information should be obtained from the manufacturer or its representatives before making the final sensor selection.Table 13 summarizes the data typically available from each sensor technology, lane coverage options, and purchase costs.The sensors discussed provide vehicle count, presence, speed, and occupancy information, with the exception of the microwave Doppler sensor, which does not offer presence or occupancy.VDSs and lidars offer vehicle classification through image processing, while microwave presence detecting radars rely on vehicle length estimates.
Roadside-mounted sensors are capable of multilane coverage, although distant lane vehicle occlusion by vehicles in the closer lanes is often an issue.In addition to the data items reported in Table 13, newer microwave radar models supply per-lane data indicative of volume, occupancy, speed, headway, and gap and per-vehicle records of volume, speed, class, range, lane assignment, and vehicle direction.
Present-day sensor technologies and models support existing traffic management functions, such as real-time operations, planning, and provision of traveler information.Some may not provide the data required for a specific application or perform as needed under the inclement weather, road configurations, lighting, flow rates, or vehicle mix anticipated at the installation site.
Certain technologies, such as VDSs, presence detecting radars, and lidars, continue to evolve by adding capabilities that identify pedestrians and bicyclists, measure multiple traffic parameters, provide individual vehicle records, track vehicles, improve resolution, operate from solar energy, simplify installation and operation, or reduce susceptibility to factors that once limited their performance.Sensors being considered for a first-time application should be evaluated under the anticipated operating conditions before making large-scale purchases.
Inductive loops and video seem to be the dominant technologies deployed for signalized intersection control, although presence detecting radars are developing capabilities that may make them appealing for this application.Magnetometers also appear in signal control, vehicle count, ramp metering, parking, and limited-access highway data collection applications.
Several areas are ripe for future research.The first is to seek methods to extract insights from the sensor data [113].This includes harvesting information that is predictive of future events to enable a controller or other management asset to initiate action that ameliorates any known negative consequences of a forecast event.A second is to incorporate efficient data fusion algorithms into edge processing when multiple sensors are present in the same device.Such algorithms have the potential to support traffic management personnel in implementing effective decisions in a timely manner.Finally, AI-based VDSs require advances to their capabilities to assign speed and class to individual objects and provide an object list that records an object's trajectory from the time it enters the sensor's field of view until it leaves [60].

Acknowledgment
The author is indebted to Pete Mills, the FHWA's contracting officer's technical representative, who provided encouragement and continued funding for the original Detection Technology for Intelligent Vehicle-Highway Systems Program, and to Don Savitt, of Hughes Aircraft, who was instrumental in selecting the author as the program manager and principal investigator for the program.I want to thank the manufacturers and vendors of the sensors whose images appear in this article for their permission to use them and for their continued technical support in assisting me in understanding the sometimes complex nature of their products' operation.Finally, words of appreciation are due to the reviewers of this article, who provided valuable insights that improved its quality.

Figure 1 (
a) shows a typical pattern of loops that might be found at a signalized intersection, while Figure 1(b) displays a speed trap arrangement of loop pairs (i.e., two loops in a lane spaced at a known distance apart) that measures vehicle speed in each lane of a limited-access highway.ILD configurations include 1.5 × 1.5 m 2 (5 × 5 ft 2 ) or 1.8 × 1.8 m 2 (6 × 6 ft 2 ) square loops, 1.8-m (6-ft)-diameter round loops, and rectangular patterns of 1.8-m (6-ft) width and variable lengths.Quadrupole-loop configurations, which divide the cross-lane loop dimension in half, enhance motorcycle and bicycle detection (by increasing the electromagnetic field strength in the center of the lane) and eliminate adjacent lane detection in inductive loops designed for high sensitivity.The increased field strength of the quadrupole loop is due to the doubling of the number of windings at the lane center.

FIG 1
FIG 1 Inductive loop configurations at a signalized intersection and on a limited-access highway (typical).(a) Inductive loops installed for signal control.(b) Inductive loops in a speed trap configuration on a California, USA, freeway.The 1.8 × 1.8 m 2 (6 × 6 ft 2 ) loops appear as a lighter color than the roadway.

2 FIG 2
FIG 2 I-Loop Duo (two-channel) card specifications for detecting, classifying, and reidentifying vehicles.(Source: CLR Analytics and Diamond Traffic Products; used with permission.)TEES: Transportation Electrical Equipment Specifications; FHWA: U.S. Federal Highway Administration; HPMS: Highway Performance Monitoring System; NEMA: National Electrical Manufacturers Association.

FIG 3
FIG 3 Magnetometer sensors provide traffic flow data, such as presence, passage, count, and lane occupancy.(a) A Sensys Networks flush-mounted magnetometer.(Source: Sensys Networks; used with permission.)(b) An M-GAGE traffic node.(Source: Banner Engineering; used with permission.)

Range to 76
m (250 ft) for up to 12 lanes.Per-lane data: volume, occupancy, speed.Per-vehicle data: volume, occupancy, gap, average speed, headway, 85th percentile speed, up to 15 speed bins, and vehicle class (up to 8).Range 6 m to 150 m (20-490 ft) for single-lane use.Speed from 4 to 300 km/h (2-186 mi/h).MOVA compatible.Provides approaching and receding vehicle detection and stationary and queuing traffic detection.Range 1.8-76.2m (6-250 ft) for up to 22 lanes.Per-lane data: volume, average speed, occupancy, vehicle class (up to 8), 85th percentile speed, average headway, average gap, up to 15 speed bins, and direction counts.Per-vehicle data: speed, length, class, lane assignment, and range.Stop bar detection to 99 m (325 ft) or stop bar and advance detection to 183 m (600 ft).Tracks vehicle speed, height, angle, and distance.Contains an integrated camera to monitor traffic and verify detections.

Figure 10 (
Figure 10(a) illustrates the constant-frequency waveform transmitted by a continuous-wave Doppler microwave sensor.The constant-frequency signal (with respect to time) allows vehicle speed measurement using the Doppler principle.Vehicle speed S is proportional to the frequency change fD between the transmitted and received signals, where fD is given by

Five
-channel, multizone sensor counts vehicles, measures speed, classifies vehicles by length (3-5 classes), detects vehicle presence and wrong-way drivers, and provides queue detection when vehicle comes to standstill under the sensor for >6 s.

FIG 12
FIG 12 Lidar sensors.(a) A LaserScan 615.(Source: OSI LaserScan; used with permission.)(b) A SICK TIC102.(Source: SICK; used with permission.)(c) A SICK TIC501.(Source: SICK; used with permission.) Lawrence A. Klein (larry@laklein.com)earned his Ph.D. degree in electrical engineering from New York University in 1973.He is a consultant with Klein & Associates, Santa Ana, CA 92706 USA.His research interests include highway and city street traffic monitoring, freeway operations, traffic signal systems and adaptive traffic signal-control algorithms, intelligent transportation systems, connected and automated vehicles, and sensor and data fusion.His books include Traffic Flow Sensors: Technologies, Operating Principles, and Archetypes (2020); Sensor and Data Fusion for Intelligent Transportation Systems (2019); ITS Sensors and Architectures for Traffic Management and Connected Vehicles (2018); Traffic Detector Handbook, 3rd Ed. (2006); and Sensor Technologies and Data Requirements for ITS (2001).He is a member of the National Cooperative Highway Research Program 03-145 panel to develop guidance for the National Traffic Sensor System Evaluation Program and the 08-157 panel to determine best practices for the data fusion of probe and point detector data.He is a Life Senior Member of IEEE.

Table 3 .
The inductive loop typical output data, installation location, advantages, and limitations.licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Table 4 .
The Model 702 microloop accuracies for vehicle count, speed, and classification.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Table 5 .
The magnetic sensor typical output data, installation location, advantages, and limitations.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Table 6 .
The VDS typical output data, camera installation location, advantages, and limitations.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Table 8 .
The microwave doppler sensor typical output data, installation location, advantages, and limitations.

Table 9 .
The PIR sensor typical output data, installation location, advantages, and limitations.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
Monitors up to five lanes.Standard output message provides per-lane measurements of vehicle volume, lane occupancy, and average speed for a selectable update period (1 to 220 s).A bit serial vehicle presence relay message or optoisolated dry contact vehicle presence relay signal is available.

Table 12 .
The ultrasonic sensor typical output data, installation location, advantages, and limitations.

Table 11 .
The passive acoustic sensor typical output data, installation location, advantages, and limitations.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.