Edge AI-Based Smart Intersection and Its Application for Traffic Signal Coordination: A Case Study in Pyeongtaek City, South Korea

.


Introduction
As the social costs of trafc congestion steadily increase, there has been a growing interest in optimizing the trafc signal controls in urban areas.Te adaptive control [1][2][3] is the one of the most well-known methods for optimizing the signal controls on road networks; however, it has the limitations in practical use regarding the stability of data collection and the feasibility of real-time computation [4].For this reason, many cities still operate pretimed control of which signal timings are calculated based on the annual average daily trafc (AADT) statistics.Accordingly, an alternative called "Smart Intersections" has been introduced recently [5,6], which is a new intelligent transportation system (ITS) solution integrating the trafc monitoring, optimal signal control, and even trafc safety.Smart intersections apply the artifcial intelligence (AI) technique to analyze video data collected from the trafc monitoring closed-circuit televisions (CCTVs) and extract the useful trafc data and utilize the processed data for trafc signal optimizations and pedestrian safety controls, etc.
Tere are several advantages of smart intersections, as they make use of CCTV video data.First of all, smart intersections are cost-efective since they do not require the road works for the construction or maintenance, unlike the ground-embedded loop detectors.Moreover, unlike other conventional trafc sensors, smart intersections can provide both point-and section-based information.In addition, smart intersections are expected to have a great potential for the signal control optimization because they can provide the contextual information, such as vehicle type classifcation, queue length, or turning ratio.
Ideally, smart intersection is an all-in-one solution for realtime intersection management that integrates multiple functions for trafc monitoring and control.However, there is a critical drawback regarding the data transmission and processing.In the current ITS system, the collected CCTV video data are practically transmitted to the ITS center and processed in a high-performance computer.Tis is because it is too heavy to operate the AI video analysis model on the on-site equipment, such as roadside unit (RSU).Accordingly, it inevitably causes at least a few seconds or minutes delays.Another issue regards their application for signal controls.Even if smart intersections have enabled to acquire more abundant trafc data, there are few cases leveraging smart intersection data.Recently, several data-driven signal control methods based on AI have been proposed [7][8][9]; however, these techniques are not matured to be practically implemented yet.Moreover, many of these initiatives require the establishment of extra equipment for collecting additional data, or some are not compatible with existing legacy signal control systems.
Tus, our goal is to construct edge AI-based smart intersections utilizing AI optimization techniques and to provide their application for trafc signal coordination.To this end, we frst install smart intersections (see Figure 1) on three consecutive intersections of Route 45 in Pyeongtaek city.Ten, the video images collected from CCTVs are analyzed on the edge devices by applying the edge AI video analysis model to extract the meaningful trafc data in real time.For the edge AI model, we compress the AI video analysis model into a small-sized one and optimize it to be well operated in the on-site edge device.Next, we provide a case study of trafc signal coordination as an application of the installed smart intersections.Te purpose of this case study is to verify the efectiveness of smart intersections on signal controls before their implementation on real roads.Tus, the experiment is conducted on a simulated environment confgured identically to the study site.Moreover, we complement some constraint conditions on signal timing variables in order to consider the compatibility with the current legacy signal control system.Te rest of the paper is constructed as follows.In the following section, the backgrounds of this research are provided.Ten, the details for constructing smart intersections and the methodology are shown.Finally, the conclusion is proposed with the experimental results.

Related Works
2.1.Smart Intersections.Smart intersections are newly proposed ITS solutions in recent years which aim to optimize trafc monitoring and control by applying AI techniques.At frst, smart intersections collect real-time trafc information by analyzing videos from trafc monitoring CCTVs with the computer vision (CV) methods.Tey detect specifc objects in the image (detection), classify the detected objects into several classes (classifcation), and track the movements of the objects (tracking).Starting with the frst application of applying deep learning to the image processing in 2012 [10], the video analysis has been greatly matured with the improvement of deep learning techniques.
Te initial algorithms for the video analysis are basically based on the convolutional neural networks (CNNs).CNNs are specialized for detecting specifc features of the image, and they are still frequently used in the feld of image processing.Starting with Regions with CNN features (R-CNN) [11], which search only specifc areas of an image, algorithms such as Fast R-CNN [12] and Faster R-CNN [13] were proposed to improve the efciency for the computational; however, these methods still have limitation in realtime video processing.In 2016, a new algorithm called you only look once (YOLO) [14] can achieve high accuracy with minimal computation, enabling object detection and classifcation simultaneously.Furthermore, recently developed YOLO v8 (by Ultralytics in Jan. 2023) and single shot multibox detector (SSD) [15] have highly improved the video analysis techniques for smart intersections.
When it comes to the trafc data, smart intersections have several advantages over traditional sensor-based trafc data collection.Most of all, smart intersections can provide both the point-and the section-based information.For the point-based information, like the loop detector and laser scanner, smart intersections can provide the fow information for vehicles and pedestrians by setting up a virtual line in the feld of view (FoV) and counting the number of objects that cross the line.For the section-based information, for example, they can measure the queue length by recognizing the stopping vehicles in FoV, like radar and lidar.However, smart intersections, in particular, can estimate the space-mean speed by measuring travel times of the traversing vehicles since smart intersections can recognize the contextualized information.For example, they can classify the types of objects into normal vehicle, bus, truck, motorcycle, pedestrian, and even emergency vehicle and personal mobility (PM), unlike radar or lidar.Finally, smart intersections can provide individual vehicle's trajectories within FoV, which is the most powerful feature.Accordingly, for example, they can measure the turning ratios without installing additional road sensors.
However, there is a critical limitation on the current smart intersection system.As heavy-sized AI video analysis models cannot be operated in the on-site equipment, the current system transmits the obtained video to a highperformance server in ITS center, which causes at least a few seconds/minutes delays (approximately 1,000∼7,000 ms at least in practice) (see Figure 2).Moreover, the existing system cannot be operated when the communication network is disconnected or where the network is not installed.To overcome this limitation, the potential use of edge AI (or called lightweight AI or on-device AI) techniques have newly been considered.It is expected that edge AI-based 2 Journal of Advanced Transportation smart intersections enable to operate the lightweighted AI model within on-site equipment (edge device) and process the collected video data in real time.Besides, as the edge system only transmits processed message data (text data) rather than full-size video, it can reduce the cost of network communication and comply with the general data protection regulation (GDPR).Moreover, it saves storing cost since it is not necessary to store all the raw video data.

Trafc Signal Coordination.
Te signal coordination usually refers to the problem that controls the ofsets of the intersection in a corridor to maximize the progression of the trafc fow for the mainstream.In general, the coordination methods aim to maximize the bandwidth which is the range of time in which a vehicle entering an upstream intersection can pass through a downstream intersection without stopping.MAXBAND [16] is the frst study which proposes the bandwidth maximization as for the signal coordination.In this study, the optimal ofset values are calculated by mixed integer linear programming (MILP) to maximize the twoway progression along the corridor.On the other hand, MULTIBAND [17] complements the relaxation condition on the feasible region of the solution to overcome the limitation of MAXBAND in which the bandwidths for each intersection are symmetrically constant.It contributes to optimize the signal coordination by considering the capacity and trafc volume of individual intersection.In addition, AM-BAND [18] suggests the asymmetric bandwidth by relaxing the existing constraints that the bandwidth is determined symmetrically from the baseline.
Unlike the conventional methodologies that the bandwidth has been determined centered at the mainstream of the corridor, recent studies consider the turning fows from the minor stream as well.In particular, the LT2 model [19] maximizes the bandwidth for both the mainstream and the side-street left-turning trafc fows to mitigate the congestion of the side-street which is hardly considered in conventional methodologies.In addition, LT2 provides a detailed modeling for the queue clearance time at the downstream intersection by considering the trafc volume and signal control variables for the upstream intersection.Nonetheless, similar to the previous methods, LT2 also assumes the uniform distribution for the trafc generation based on statistical trafc volume aggregated in a large range of time window, which may not be appropriate to the actual trafc.

Problem Statement
3.1.Study Site.For the case study, we target the problem of trafc signal coordination in Route 45 of Pyeongtaek city, South Korea.We construct smart intersections in this target study area aiming to improve the signal coordination.Specifcally, the spatial range includes three consecutive intersections of Route 45 in Pyeongtaek city, South Korea, as shown in Figure 1.Tis section is a major intercity arterial that connects the central Pyeongtaek area (North) to Asan city (South).Tis section also has a number of trafc demands not only for the commuting vehicles but also for the heavy vehicles, such as cargo trucks.As the majority of trafc demands travels from north to south, the signal coordination is set to accord with the same direction.However, this coordination setting is not efective to the nonpeak hour trafc demands since it yields unnecessary delays to the opposite direction (South ⟶ North) or turning fows.Tus, the temporal range of this study is confgured as 13:00∼16:00 when public petitions are frequently registered.

Current Status and Gaps.
We frst collect 24-hour trafc data on 18 May 2022 (Wednesday) after installing smart intersections to identify the current status and to investigate research gaps.Te analyzed results are shown in Figure 3. Te top of Figure 3 shows the changes in trafc volumes, while the bottom shows the turning fows during the time for TOD PLAN #2 (08:30∼16:00) which includes the target temporal range (13:00∼16:00).Te results show that the study site has a high level of trafc demands during the peak hours, and the demand of the mainstream (North ⟶ South) is especially high.In addition, the majority of trafc demands at Pyeonggung-samgeori (3-way intersection in the middle) travel along with the mainstream; however, 11% merges to the opposite direction of the mainstream from the minor roads (Anjeong-ro).
Te signal control for this study site is operated by pretimed TOD calculated based on AADT, and the signal information including phase design and minimum green time is shown in (Table 1).Te overall TOD plans are given in Table 2.It is seen that each intersection has four TOD plans and shares the common schedule and cycle time since all the intersections belong to one subarea (SA).Te time-space diagram for TOD PLAN #2 is plotted in ((a)), and it can be seen that the signal coordination is set to accord with the direction for the mainstream (North ⟶ South).Accordingly, the Journal of Advanced Transportation majority of the mainstream fows can pass through the area without stopping.However, this coordination setting is not efective to the nonpeak hour trafc demands since the TOD plan is based on the aggregated statistical historical trafc data.For example, although the number of trafc fow for the opposite direction (Path 2 in Figure 4) increases up to 70% of that of the mainstream during the target time range 13:00∼16:00, it fails to coordinate and the platoon is cut of at Pyeonggungsamgeori.In addition, the left-turning fow at Pyeonggungsamgeori merging into the opposite direction of the mainstream (Path 4 in Figure 4) increases up to 35% of that of the mainstream; however, the majority fails to coordinate, and the platoon is cut of at Pyeonggung-sageori (4-way intersection at north).In the meantime, even if the left-turning fow from the eastern approach of the Pyeonggungsageori (Path 5 in Figure 4) decreases below 1% of that of the mainstream, it unnecessarily coordinates the signal so that the corresponding trafc fow can pass through the area without stopping.In conclusion, the existing signal coordination is only centered at the mainstream that results in coordination failure for the opposite direction and leftturning trafc demands in spite of their demand levels are not low.

Construction of Edge AI-Based Smart Intersections.
A key clue for resolving the coordination failure of the study site is to acquire real-time trafc fow information for each   approaching link and recalculate ofsets according to these data.Hence, we install CCTV cameras on the downstream of each approaching link to capture the turning fows and queue information.Additionally, we install edge devices on each intersection to process the collected video images from the CCTV cameras using the AI video analysis model in real time.Te components are described as in (Figure 5).Next, we have the optimized lightweight AI video analysis model via NetsPresso (AI optimization solution provided by Nota AI Inc. (https://netspresso.ai/)) (AI optimization platform developed by Nota Inc.).Te mechanism of NetsPresso is as follows: at frst, we have a pretrained object detection model using labeled intersection image data.In this study, we use YOLOX as a backbone which is a high-performance one-stage model employing a decoupled head [20], and the model is fnetuned for each camera's FoV.Ten, the importance for each flter of the CNN is measured using the structured pruning technique [21].Te importance is defned by the L 2 -norm for the weight parameters of the CNN flter.Te less important flters are removed to compress the model size.Tis process is repeated until the model size is smaller than the target size.Besides, for object tracking, we use the discriminative correlation flter (CDF)-based visual tracker [22].Finally, the compressed model is converted and packaged to be mounted on the edge devices installed in the study site [20,22].(Te specifcation of the edge device is shown in Table 3).
Te region of interest (RoI) for the object detection is set as in Figure 6.At frst, the range is set to be the maximum distance in the camera's FoV where the object's type is distinguishable, and the region is divided by each lane.Ten, unlike the existing approaches for smart intersections, we additionally include the part of the upstream of opposite direction in the RoI to measure both infows and outfows.Figure 6(b) shows the result of inference of the AI model, and it can be seen that the objects in both downstream and upstream are detected and classifed into each vehicle type.
From the video analysis, we collect the trafc data: trafc volume and the number of queueing vehicles by lane and by vehicle type, average speed of each lane (space-mean speed in each RoI).At frst, the objects are classifed into three categories: car, bus, and truck.Ten, the trafc volume is measured by setting up a virtual line and counting the number of vehicles crossing the line.Te queue information is measured by counting the number of vehicles moving at less than 5 km/h for a certain period.Furthermore, the travel time of each vehicle passing through the RoI range is measured, and the space-mean speed for the RoI is estimated by harmonically averaging the travel time.

Trafc Signal Coordination Method.
As this study aims to treat the signal coordination failure problem in nonpeak hours occurring unnecessary delays of the side-streets with relatively high demands, we propose to utilize the LT2 model  6 Journal of Advanced Transportation to coordinate the multidirectional trafc fows.We adopt the basic structure of the LT2 model as the backbone; however, we partially adjust the model to use the real-time trafc data collected from the smart intersection.Besides, we derive the conditions for its application in the legacy signal control system and add them into the constraints.First, the trafc volumes of each lane collected from the smart intersection are aggregated according to the turning directions.Ten, the aggregated directional fows are used as a major input variable for the model.Second, we adopt the objective function which is the jointly maximization of the bidirectional bandwidths and the side-street left-turning bandwidths, as in equation ( 1): where Te key constraint conditions of LT2 are as follows (directly referred from [19]): for i � 1, 2, . . ., n − 1,

Signal info
Real-time scope   Journal of Advanced Transportation Equation ( 2) is to utilize the constraints of MULTI-BAND, which is fundamentally required to achieve an equation coordinated bandwidth model.Equation ( 3) is to relax the existing constraints on the bandwidth by modeling the queue clearance time with observed upstream infows.Equation ( 4) is to describe the relationship between the bandwidth of side-street turning fow and the signal phases.
Tird, we additionally consider the following constraint conditions regarding the legacy signal control system: preservation of cycle time and preservation of green split in each TOD.In the current legacy system of South Korea, changing cycle time only for a few intersections in one SA group is not allowed.Likewise, changes of green splits are not easy to be allowed due to the stability issue so that we set it as a hard constraint.Instead, simply changing the ofset values is relatively easy to be applied in the legacy system, as it only changes the starting time of the existing TOD plans.Other crucial constraints, such as preservation of phase design, phase sequence, ring design, are also considered.
Finally, we interpret the output of the LT2 model as the ofset values of each intersection, as the bandwidth which is the output of LT2 model can be simplifed to an equation by the ofsets according to the above constraints.Te description for other variables is summarized in Table 4.

Experimental Design.
We set up a virtual environment using AIMSUN, a microscopic trafc simulation tool to evaluate the performance of the proposed model in the target area.To replicate the installed smart intersections, the arterial links are divided into upstream, midstream, and downstream sections based on the RoI range of the camera.Te upstream and downstream sections represent the areas within the RoI where the trafc data can be extracted, and the midstream is a blind section so that the trafc data in this section are not collected.
Next, the collected real-time trafc data are aggregated at intervals identical to the signal cycle length, constituting one data unit.Trafc variables, such as in/outfow and turning ratios, are derived within the unit.Ten, the outfow and infow are embedded into the downstream and upstream links, and the turning ratios are embedded into each node.Tis approach allows to create a virtual trafc environment that is similar to the actual study site.To relieve the data fuctuation, these units are aggregated in 15 minutes and it confgures the demand scenario.Te model performances are evaluated in the scenarios with the same random seed, and the fnal result is derived by averaging the results across the scenarios of 10 diferent random seeds.
To measure the efects of the proposed model, we compare the performance with other well-known signal coordination methods, such as MULTIBAND, PASSER2, and the existing TOD plan.For a fair comparison, we maintain the same constraint conditions as the legacy system, such as cycle length, phase order, and green splits, but it only controls the ofsets.Additionally, this approach enables to solely evaluate the impact of changes in the bandwidth to the trafc fows, excluding other factors.
For the evaluation, we employ the average number of stops as the primary measure of efectiveness (MoE) since this study aims to maximize the bandwidth of bidirectional and turning trafc fows through ofset control.In addition, the average travel time and the average delay serve as secondary evaluation metrics to measure the network performance.Te average number of stops is normalized by the travel distance to obtain the average number of stops per unit length (#/km) since each vehicle has a diferent route.Similarly, the other two time-related metrics are also normalized as the average travel time per unit travel distance (sec/km) and the average delay per unit travel distance (sec/ km), respectively.
Te explicit forms of these metrics are as follows: for all vehicles entering the network, veh i (i � 1, 2, . . ., N), the vehicles that traverse each route P j are denoted by , and the travel distance of P j is denoted by L(j).Subsequently, the average number of stops throughout the network and the average number of stops for each P j are denoted by s and s j , respectively, and they can be calculated based on the stop time of each vehicle i, denoted by s(i).
Similarly, if we denote the travel time of i by t(i) and the delay time by d(i), then the overall average travel time in the network t, average delay d, the average travel time t j , and the average delay d j for each route P j are calculated by

Experimental Result.
Te optimized AI video analysis model is applied on CCTV videos to extract the real-time trafc data for the study site.Te performance of the AI model optimization is summarized in Table 5.First, the model is signifcantly compressed of which size is decreased by 97.8% compared to the original model.Tis means that the compressed model takes only 2.2 Mb if the original takes 100 Mb because a number of weight parameters are eliminated.Second, the optimized model can process incoming video data in near real time.In general, inference speed measures the performance of AI model lightweighting, and 30 FPS is considered as "realtime."On the installed edge device, the proposed model shows 29.49FPS which is near real time.Finally, the model maintains a similar level of accuracy despite the compression.In general, accuracy tends to decrease when the parameters are eliminated through model compression.However, the size of the model can be reduced to a level that maintains the accuracy by selectively eliminating less-contributing parameters.To test accuracy, the model is trained using 8,824 collected image frames including cars, buses, trucks, motorcycles, and pedestrians.Ten, the model is validated with 100 unseen image frames of which ground truth is manually counted.Next, we utilize the real-time trafc data extracted from the smart intersections as input variables in equations ( 1)-( 4) to calculate the optimal ofset for each intersection.We apply mixed-integer nonlinear programming (MINLP) Out/in-bound arterial progression through bandwidth (cycles) on section i al i (al i ) Weight for out/in-bound progression cross band on section i bl i (bl i ) Out/in-bound side-street left turn green bandwidth (cycles) on section i v mt i (v mt i ) Out/in-bound arterial through volume (veh/hr) at Out/in-bound side-street left-turn volume (veh/hr) at S i S i (S i ) Saturation fow on section i r i (r i ) Out/in-bound red time at S i (cycles) Time from right side of red at S i to S i+1 outbound (S i+1 to S i inbound) (cycles) Travel time from S i to S i+1 outbound (S i+1 to S i inbound) (cycles) Time from center of an out/in-bound red at S i to the center of a particular out/ in-bound red at S i+1 ∆ i Time from center of r i to nearest center of r i (cycles) Out/in-bound left-turn green time on cross street at S i (cycles) dls i (dls i ) Time from end of out/in-bound arterial through green phase to start of out/ in-bound side-street left turn green phase at S i (cycles) Uniform arrival rate (veh/sec) for vehicles departing during arterial through green from S i to S i+1 outbound (S i+1 to S i inbound) Uniform arrival rate (veh/sec) for vehicles departing during side-street left-turn green from S i to S i+1 outbound (S i+1 to S i inbound) Uniform arrival rate (veh/sec) for vehicles turning right from cross street during red time of coordinated phase from S i to S i+1 outbound (S i+1 to S i inbound) Outin-bound time diference of red starts of coordinated phase between S i (S i+1 ) and Out/in-bound arterial saturation through fow headway (sec/veh) at S i to solve the optimization problem in equation ( 1) that involves integer variables using CPLEX (version 12.3) API provided by IBM.Te calculated optimal solutions are then applied as the ofset value of each intersection into the AIMSUN environment.
For a detailed evaluation, we analyze the MoEs not only for the entire network but also for the selected 5 specifc routes, as illustrated in Figure 4. Te frst route, named by Path 1, corresponds to the major traveling direction on the mainstream which has the highest level of trafc volume.On the other hand, Path 2 is selected by the opposite direction on the mainstream to evaluate the efect of maximizing bidirectional bandwidth.Moreover, we also consider Path 3 and Path 4 which have relatively high trafc demands among the minor streams to measure the coordination efects on the side-street left-turning fows.In addition, Path 5 is also included of which signal is coordinated to the mainstream despite the trafc demand is signifcantly low.
Te numerical results are summarized in Figure 7.Most of all, it is found that LT2 improves network efciencies in every MoE.Compared to the existing TOD, the average number of stops is decreased from 1.04 to 0.96, indicating approximately 7.69% improvement.Similarly, the average delay and travel time are improved by approximately 6.2% and 2.92%, respectively.PASSER2 and MULTIBAND also improve the network performances compared to TOD.
However, upon examining the results for individual paths, it becomes evident that LT2 shows better performances.Specifcally, both MULTIBAND and LT2 similarly exhibit the improvement on the mainstream, Path 1, while PASSER2 shows the worst performance.On the other hand, for the two major side-street left-turning fows, Path 3 and Path 4, it is remarkable that LT2 improves the performance than MULTIBAND.It implies that the LT2 reduces unnecessary delays of the sidestreets with relatively high demands.Additionally, it can be seen that the existing TOD unnecessarily yields the most efective signal coordination to Path 5 which has the lowest demand.
Te changes of bandwidth can be observed in Figure 8, and it corresponds with the numerical results analyzed in Figure 7.In the outbound direction, the LT2 and MUL-TIBAND models present an expanded bandwidth ⓐ for the major fow, surpassing the TOD and PASSER2 models.Terefore, they allow a larger number of vehicles to pass through the corridor (Path 1) without stopping.In addition, the left-turning fows for Paths 3, 4, and 5 are allocated to ⓑ, ⓒ, and ⓔ, respectively.It is observed that TOD inefciently assigns wider bandwidth to ⓔ, yet relatively narrower bandwidth to ⓑ.In contrast, the LT2 model efectively distributes sufcient bandwidths ⓑ and ⓒ to Paths 3 and 4, which have relatively high demands, and manages to efciently accommodate Path 5 as well, unlike the MULTI-BAND, which fails to secure bandwidth ⓔ.

Discussion
In this study, each of the four signal coordination models requires distinct spatial and temporal resolution for trafc data.Te existing TOD, based on AADT statistics with low temporal resolution, shows signifcant limitations in adapting to fuctuating trafc demands.To improve this, real-time trafc fow data collected by loop detectors installed in straight lanes of the mainstream conventionally facilitate the signal coordination algorithms, such as PASSER2 and MULTIBAND.Tese conventional signal coordination algorithms improve the network efciency centered at the mainstream, as shown in Figure 7.However, there still have been signal coordination failures on irregular travel demands during nonpeak hours occurring unnecessary delays of the side-streets with relatively high demands (e.g., degradation of MULTIBAND for Paths 3 and 4).
Te state-of-the-art signal coordination methods, including LT2, propose novel methods to coordinate the multidirectional trafc fows in order to mitigate the congestion on the side-streets with relatively high demands.Although these algorithms demonstrate signifcant improvement in their simulation-based experiments, they would encounter some challenges with regard to the practical implementation.Tese methods require highresolution real-time trafc data for the turning trafc fows of each intersection, such as turning ratios and queueing vehicle numbers, in order to calculate the accurate values for the signal timings.
In this aspect, the edge AI-based smart intersection proposed in this study highlights the potential use of these novel signal coordination methods by serving highresolution trafc data in real time.Taking the advantages of using CCTVs and AI, the edge AI-based smart intersection provides abundant trafc data of point/sectionbased information, and even contextualized information, unlike the other traditional VDS.Accordingly, this study provides an application of improving signal coordination using real-time trafc data collected from edge AI-based smart intersections.By leveraging these data, the experimental results indicate that LT2 alleviates the coordination failure problem for nonpeak hour demands in the study site.Journal of Advanced Transportation

Conclusion
Te goal of this study is to construct edge AI-based smart intersections utilizing AI optimization techniques and to provide their application for trafc signal coordination.To this end, we install smart intersections on three consecutive intersections of Route 45 in Pyeongtaek city, South Korea, and collect the real-time trafc data by applying the edge AI video analysis model.Te model compressed and optimized via NetsPresso maintains a similar level of accuracy (93.64%), even if the size is reduced by 97.8% compared to the original.Next, we utilize a LT2 model to treat the coordination failure problem in nonpeak hours occurring unnecessary delays of the side-streets with relatively high demands.We complement some constraint conditions in order to consider the compatibility with the current legacy signal control system.Te experiment is conducted on the virtual environment of which geometry and trafc demand are confgured based on the features of the installed smart intersections.Te numerical results conclude that the calculated optimal ofsets calculated by the LT2 model efectively manage bandwidths for multidirectional fows based on the real-time trafc demands collected from the edge AIbased smart intersections.
Te main contribution of this research is that it introduces an edge AI-based smart intersection.Although smart intersections have been prevalent in many cities, there are a few drawbacks in their operations.In this regard, this study demonstrates the efectiveness of edge AI-based smart intersections by extracting real-time trafc data from CCTV video data, even on low-powered edge devices, with high accuracy.Furthermore, this study explores the application of edge AI-based smart intersections to a practical signal coordination problem using a state-of-the-art algorithm that requires high-resolution real-time trafc data for all turning trafc fows of each intersection.Tis research serves as a preliminary study to validate the efectiveness of edge AI-based smart intersections in signal coordination before conducting on-site tests.Te primary future plan is to carry out experiments on actual roads rather than in a simulated environment.Subsequently, the performance of the proposed method will be assessed using real trafc data.Furthermore, we plan to explore additional signal control variables, including green splits or cycle time, as part of our efort to revise the legacy system.Moreover, future studies will involve the development of an enhanced model, leveraging a broad spectrum of trafc data obtained from edge AI-based smart intersections.

Figure 1 :
Figure 1: Layout of the study site.

Figure 6 :
Figure 6: Result of object detection of smart intersection.(a) Region of interest (RoI) setting.(b) AI model inference for object detection.

Table 1 :
Signal phases (upward arrow points to north).

Table 4 :
Key model parameters and description.Weight for out/in-bound arterial progression through band on section i bt i (bt i )

Table 5 :
Performance of edge AI video analysis model.