1. Introduction
According to the global status report on road safety [
1], 1.35 million road accidents occur globally in a year, of which developing countries account for around 90% of all casualties. Pedestrians, cyclists and motorcyclists constitute 54% of all road accident-related deaths worldwide, not only resulting in personal losses, but also national economic losses. Furthermore, due to rapid urbanization and the growing motorization rate, the rate of crashes has increased exceptionally, with an average of 1230 accidents and 414 deaths every day, and nearly 51 accidents and 17 deaths every hour [
2]. This renders road safety a very crucial issue, especially in developing countries like India. One of the critical reasons for this is the diverse traffic conditions in India, with high-speed vehicles of differing characteristics sharing the same road space, creating a heterogeneous traffic condition, which makes road safety analysis complex. Moreover, intersections are particularly dangerous because of risky driving maneuvers, such as sudden shifts in vehicle speed, unforeseen lane changes and lane indiscipline [
3].
For the safety analysis of intersections, traditional techniques have been used in India. However, one of the most substantial concerns with using the conventional techniques is their reliance on the occurrence of serious accidents before any hazardous site or situation can be identified and corrected [
4]. Numerous studies have extensively discussed the drawbacks of relying solely on conventional traffic accident data [
5,
6,
7].Therefore, surrogate safety measures have been suggested as a viable alternative to traditional accident data to address these issues. These measures are explanatory measures that can be easily recorded or gathered that relate to road networks and vehicle behavior. They are independent of direct crash data and have grown in importance in recent studies on traffic safety [
8]. The main goal of these measures is to estimate the predicted frequency of crashes and injuries by evaluating non-crash incidents and to obtain a deeper understanding of the numerous causes and mechanisms that contribute to accident occurrence [
9]. The key benefit of using surrogate safety measures is that they occur more often than crashes and have a proactive approach to analyzing traffic conflict, with a smaller observation period and better statistical results. The primary assumption underlying the entire surrogate safety measures theory is that traffic incidents can be classified into stages according to their severity level. Historically, traffic safety severity has been commonly presented as ‘proximity to an accident’. The adoption of the Vision Zero paradigm in traffic safety ((aims to eliminate all fatalities and serious injuries in traffic [
10] marks a change from an overall focus on accident prevention to a more targeted approach that focuses specifically on the subset of accidents resulting in serious injuries and causalities. To align with this emerging perspective, the concept of severity for traffic occurrences should be updated and defined as “proximity to a serious injury”, instead of “proximity to an accident” [
11]. However, in this definition, it is unclear how severity can be assessed using objective and accessible indicators, which is a requirement for surrogate measures to be helpful rather than merely a theoretical idea [
6]. While there is empirical evidence suggesting the credibility and accuracy of how human observers perceive dangers in traffic situations (which is explained in detail in the following section), it is critical to recognize that using human observers as a measuring tool includes several drawbacks. When processing large amounts of traffic data, using humans can be costly and often unfeasible. On the other hand, as automated data collecting tools are becoming increasingly common in the analysis of surrogate safety, it is necessary to represent the theoretical severity using a combination of objective metrics and the indicators derived from the data that these tools generate. Therefore, the main hypothesis of this research is to conduct a comprehensive exploration into how human assessments of traffic situation severity can be translated into objective safety indicators under heterogenous traffic conditions. This involves creating a decision-making system that assesses the severity of two traffic situations and determining which of the objective indicators most closely aligns with the human decision when faced with the same choice.
In order to conduct this study, videographic data were collected during daytime hours, under normal weather conditions in the summer season at six uncontrolled intersections situated at Tier II cities of India. These conditions provided drivers with unobstructed visibility and a realistic portrayal of road safety in optimal circumstances. As visibility is one of the crucial factors in road safety [
12] and it encompasses various aspects that can significantly impact driving conditions—especially at night times where drivers need to mitigate visibility-related risks [
13]—it is crucial for drivers to adjust their behavior and use the appropriate precautions, depending on the specific conditions [
14]. This may include reducing speed, using headlights or fog lights and maintaining and being vigilant of potential hazards. Therefore, as the data collection in this study was conducted under very specific weather and environmental conditions, the results will only be applicable if the specific conditions are maintained.
This paper is organized as follows: The following section presents the literature review relevant to surrogate safety measures and human perception of safety and shows the need for this study. The third and fourth sections present the data collection and data extraction-indicator estimation processes. The fifth and sixth sections discuss the analysis of road user behavior and the relationship between the human perception of safety and objective measures using ordinal logistic regression. Lastly, the concluding section summarizes the paper and provides conclusions regarding the study, including the limitations and areas for future research.
2. Literature Review
Over the past few years, road safety monitoring has evolved, and a more proactive approach has been adopted. Earlier road safety analysis was conducted using the police crash data report, but there are well-known issues with these kinds of reports. Being a reactive process, it has significant drawbacks like underreported cases, a lack of behavior analysis and is very time-consuming [
15,
16]. Therefore, surrogate safety measures have been introduced, and a traffic conflict technique has been developed. In the traffic conflict technique, in a scenario where two or more commuters are approaching one another in time and space, a collision will eventually occur if their movements do not change, referred to as a conflict [
17]. Traffic conflict measures are typically used to identify when a traffic event transitions from a typical interaction to a conflict and, ultimately, to a collision [
3]. Most (more than 50%) crashes occur at urban intersections due to dangerous driver actions and movements [
4]. One of the reasons can be drivers from different directions occupying a single space at a junction, which reduces network capacity, delays drivers and poses a safety concern [
7].
The definition of surrogate measures includes not only parameters like Post Encroachment Time (PET), Time to Collision (TTC) or TTC-Integrated, but also traffic characteristics such as traffic volume, delay and speed [
6]. By incorporating these additional factors, the definition of surrogate measures becomes more comprehensive and inclusive and considers anything that could potentially lead to conflict on the road as a surrogate measure. The most often used safety indicators, such as Time-to-Collision [
18] and Post-Encroachment Time [
19], and their modifications only evaluate one aspect of safety [
20], i.e., proximity between two road users who may collide. This measurement only accounts for one part of the possible accident severity. The other aspect of safety, the potential repercussions of a hypothetical conflict, has been considered only by a few researchers. The Swedish traffic conflict technique is one of the few techniques that considers both the TTC at the moment where the road user detects the risk and takes evasive action and the speed of the road user at that instant to evaluate the severity of a traffic event [
21,
22]. A higher speed with a low TTC value is believed to raise the severity of the potential repercussions, exacerbating the danger. In contrast, the Dutch objective conflict technique for operation and research (DOCTOR) uses a subjective score representing the likelihood of injuries in a collision [
23,
24]. The score is determined by objective criteria such as the road user and maneuver type, controlled or uncontrolled situation and specific attributes like the presence of evasive action. Other indicators, like Delta-V and combined risk indices, have also been proposed; however, their application is constrained as they rely on specified assumptions and are only valid in certain circumstances. Svensson [
17], in a validation study of the Swedish traffic conflict technique, included a subjective measure of conflict danger. The conflicts chosen based on the danger rating showed the strongest association with accidents reported by the police. Human perception of danger is a comprehensive process that is essential for survival and decision-making to lessen the possible harm; it may inspire people to take preventative measures, stay away from dangerous circumstances or participate in protective behaviors. It is crucial to understand how people perceive danger in disciplines like risk assessment or traffic safety engineering.
One of the important questions is whether depending on human observers’ perception of danger in a traffic event is an appropriate way to determine ground truth or if it is biased. However, some empirical evidence suggests that human judgments are reliable and valid for detecting severity in traffic events. During the 1980s, which was a significant period for the advancement of traffic conflict techniques based on human observation, multiple studies were carried out to evaluate the reliability of observers. The findings showed that with the right initial training, there is good agreement between observers in terms of identifying traffic conflict [
22,
25]. Numerous calibration studies, comprising up to ten distinct traffic conflict techniques (TCT), have also confirmed a high level of agreement between observers regarding the severity-based rating of identified conflicts based on their risk level [
24,
26]. In conclusion, it seems humans have an innate ability to assess the level of danger in a traffic situation, particularly when it is based on an actual observation rather than a hypothetical scenario. Furthermore, this technique could be the most effective means of gauging the theoretical notion of “severity” as ratings derived from human judgments cannot contradict common sense [
11].
Nonetheless, human observers have certain limitations and can be used to a certain extent as a measuring tool. Using human resources to analyze large amounts of traffic data can be expensive and sometimes impractical for analyzing extensive traffic data. Also, more recently, many researchers have used automated traffic data collection and extraction tools [
15,
20,
27] and trajectory-based data for surrogate safety analysis [
28,
29,
30,
31,
32,
33]. Using micro-level behavioral data, Laureshyn et al. [
30] proposed a conceptual framework to assess traffic safety using trajectory data. Continuous profiles were calculated for a detailed understanding of the interactive process, and the continuous trajectory profiles of indicators (TTC & GT) of individual interaction of pedestrians with vehicles were studied [
32,
33,
34]. Considering that, it is important to establish the theoretical severity through the indicators available in the data generated by these tools.
This study aims to first conduct a detailed analysis of road user behavior traits in various vehicle–vehicle interactions, and then to thoroughly analyze how human assessments of the severity of traffic situations can be expressed using objective safety indicators. To achieve this, an ordinal logistic regression technique was used that compares the severity rating given by observers and investigates which objective indicators most closely align with human decision-making for the same selection.
4. Data Extraction and Indicator Estimation
To assess the overall behavior of the road users in our study, a comprehensive trajectory profile of the vehicles was required. After reviewing the literature, Kinovea (semi-automated tool) was found to be the most suitable tool for accurately tracking vehicles [
33,
35,
36,
37]. One of the key advantages of using Kinovea is that it is an open-source and free tool with a user-friendly interface. Thus, Kinovea 8.27 was used in our study to extract the linear kinematic characteristics—such as the trajectories, speed and distance—from the video data of all the intersections.
Table 2 summarizes the various indicators that were derived from these trajectories, along with their brief description.
Several safety measures that can accurately calculate the safety of a traffic event have been proposed in earlier studies. For example, time to collision (TTC) and post encroachment time (PET) are some common safety measures [
20]. These measures can be used to anticipate possible crashes and near misses, providing important information about the safety of traffic situations. Many of these measures can be derived from the trajectories of road users and a few other characteristics, like the weight and dimension of a vehicle [
30].
Only those indicators that can be derived from the trajectories of vehicles were used in this study. Some clarification regarding the choice of indicators is presented in this section. Although TTC is the most commonly used surrogate safety indicator, its applicability is somewhat constrained; it requires the presence of a collision course to calculate its value and cannot be calculated for many of the less severe scenarios, which is a drawback. Given that, in this study, we are interested in situations spanning the complete severity continuum,
(time remaining for the latest road user to arrive at the potential conflict area) was chosen over TTC as it is a more adaptable measure [
20,
30]. TTC counts the time before a potential collision, whereas
measures the time needed for the latest road user to reach the conflict area. When a crash is about to happen,
gives the same advantages as TTC, but also supplies helpful information when there is no chance of a collision. As a result, this study used
rather than TTC as a reliable safety indicator. The other measure used in this study is distance. The main benefit of using distance as a measure is the simplicity of its calculation.
There is an advantage to using straightforward universal measurements in comparison to complex ones that may be subject to misinterpretation [
38]. This study focused on two forms of distances: the shortest straight-line distance between the road users’ closest points and the distance along their predicted paths to the point of conflict (see
Figure 2).
The other surrogate measure is PET, which is defined as a time interval between the first road user’s departure and the second road user’s arrival at the point of conflict [
19], whereas gap time is used to check the temporal closeness of the vehicles [
39].
The deceleration-based measures and speed were also estimated for the vehicles involved in the interaction to illustrate the possible consequences if the crash occurs.
Table 2.
Objective Indicators.
Table 2.
Objective Indicators.
Indicator | Definition |
---|
Evasive action (EA) | Evasive action is a categorical variable (If EA is present = 1, if not EA = 0) that indicates whether any of the road users have taken any evasive action or not. |
Time until collision (T2) | The amount of time remaining until the latest road user reaches the potential collision point [30]. When on a collision course, this is equal to the TTC. |
Shortest distance (D1) | Euclidean distance (Shortest distance). |
Sum of distance (D2) | The sum of distance travelled by the road users to reach the potential conflict area. |
Post Encroachment time (PET) | The time interval between the first road user’s departure from the potential conflict area (who arrived there first) and the entry of the second road user [19]. |
Deceleration rate (DR) | Deceleration rate of the road user. In cases of acceleration or no deceleration, DR is taken to be zero [40]. |
Speed (V) | Speed of the vehicles involved in an interaction. |
Gap Time (GT) | The time interval between the second vehicle arriving in the potential conflict area after the first vehicle leaving the conflict zone when both proceed at the same speed and trajectory [39]. |
5. Behavioral Analysis
This section presents a detailed analysis of various vehicle categories’ behavior involved in an interaction. The trajectories of the vehicles were extracted from the video data collected at multiple study locations, and various indicators were estimated using these trajectories. Three indicators—namely, time until collision, gap time (GT) and speed (V)—were used to evaluate the behavior of the road users during an interaction, and the remaining indicators were used in establishing the relationship with human perception of danger. An example of an interaction between two aggressive two-wheelers is presented to provide a better understanding.
Example
This segment shows an example of a behavioral pattern where none of the vehicles took any evasive action. The speed, time until collision
and GT profiles of both vehicles were estimated using extracted trajectories.
Figure 3, given below, depicts the interaction trend between two two-wheelers at a right angle, along with the profiles of the indicators in
Figure 4.
The following interaction occurred between two two-wheelers (labelled in a green and yellow box) and is individually determined as a risky case. The entire interaction phase was initiated from the point when the vehicle (green box) observed the approaching vehicle (yellow box); at point , the second vehicle shows up in the range of assessment. Meanwhile, the vehicles managed to maintain a high speed.
The reduced until it attained the minimum value at as both the vehicles approached the possible conflict point. The GT, which permits a whole vehicle to cross through the possible conflict point, also decreased after a 0.5 s lag and attained its lowest value between and . crossed the possible conflict point just before the second vehicle reached , indicating interaction completion, whereas at , both the vehicles crossed the conflict point.
In the research sites, various categories of vehicles were observed, such as two-wheelers, cars, auto rickshaws, buses, light motor vehicles, e-rickshaws, etc. Among the total traffic observed at all sites, the major vehicle categories were two-wheelers, cars and auto rickshaws, comprising approximately 95 percent of the total traffic. Hence, only these categories were used to conduct the subsequent detailed analysis in order to obtain unbiased results.
Figure 5, below, shows sample graphs of pair wise vehicle–vehicle interactions, plotted to analyze the pattern and to determine if there is any notable difference in the indicators’ profile with the changing vehicle categories. Road user behavior was classified into two categories according to the indicators’ profile shape and the presence of evasive action. The details of both categories are as follows:
- (1)
Non-Aggressive interaction: When one or both vehicles take any evasive action, it can be categorized into receptive road user behavior (non-aggressive behavior). Upon analyzing the
and GT profiles in
Figure 5a, it is apparent that the vehicle interaction with various other vehicle categories generates a similar profile shape. The minimum values for both
and GT occur towards the middle or end of the interactive process.
The GT value starts to drop as two vehicles get closer to a potential conflict point until it hits its lowest value, which is close to zero, signalling that the two are most likely to arrive at the conflict point simultaneously. However, because of evasive action taken by the road user, the gap time started to increase again. The lowest value of GT can therefore be described as a risky point.
In
Figure 5a, when the speed profiles of different interactions were analyzed, it was observed that the speed profiles decreased over time. And, once it reached the minimum value, it started to increase again, indicating that one or both road users had taken significant measures to avoid a potential collision.
Therefore, this behavior can be classified as a non-aggressive interaction as it involves an evasive action.
- (2)
Aggressive interaction: In this type of interaction, it was observed that despite having low values for
and GT, neither of the road users took any evasive action. As can be seen in the speed graphs in
Figure 5b, the speed profiles are nearly horizontal with respect to time.
As the vehicles approach the potential conflict point, the distance between them starts to decrease until one of the vehicles crosses the conflict point. Therefore, the riskiest point occurs near the end of the interaction. Hence, this behavior can be classified as an aggressive interaction that involved no evasive action, despite having a risk of collision. Also, when observing the trends of different vehicles, the and GT profiles are different.
Distinguishing road user behavior visually based on speed profiles is a challenging task. Hence, the behavior can be identified based on the time when either
or
occurred. In non-aggressive behavior, it is noticeable that
and
occurred simultaneously (either at the end or middle of the process) for all vehicle categories. However, this was not the case for aggressive behavior. In this behavior, a clear differentiation can be perceived in the occurrence of
and
. This disparity in the occurrence time of
and
was attempted to be illustrated mathematically, as shown in equation below:
It was assumed that, if the interaction starts at time , then the two vehicles will each arrive at the potential conflict zone at time and , respectively. As a result, represents the moment when the first road user reaches the conflict zone, and represents the moment when the second road user reaches the conflict zone.
If Equation (1) is fulfilled, it can be inferred that an interaction belongs to the first type of road user behavior; otherwise, it belongs to the second type (aggressive) of road user behavior. Additionally, and represent the minimum and GT values, respectively, and δt represents the interval between these values.
As indicated in
Table 3, 1141 interactions were divided into two types of road user behaviors using Equation (1). It can be observed from
Table 3 that the vehicles most frequently involved in non-aggressive road user behavior are two-wheelers and cars; the bigger size and width of these vehicles, which can psychologically induce vehicles to take evasive action, is one of the important causes of this. On the other hand, in aggressive road user behavior, the most common vehicle interaction type is two-wheeler–two-wheeler, which shows that two-wheelers comprise the most risk-taking road users.
6. Human Perception of Severity and Objective Measures
Further analysis was conducted to develop a relationship between the observer ratings and the objective measures. The detailed study was conducted on all 1141 interactions, and the following behaviors were observed in their trajectories:
- (a)
The First vehicle accelerates and/or alters trajectory to cross a conflict point before the second vehicle arrives;
- (b)
The Second vehicle slows down, stops or changes trajectory to avoid a collision;
- (c)
Both vehicles take evasive action, and one of the vehicles takes the lead and crosses a possible conflict zone before the second vehicle;
- (d)
None of the vehicles take any evasive action.
A trained observer technique was then used to rate the severity of the interactions using the above behavior patterns. Five trained observers, all from different cities of India, were first informed of the purpose of the study, along with the fact that the likelihood of being involved in a collision depends on both how close the incident is to becoming a crash and what steps the involved vehicle or vehicles take to avoid a collision before they arrive at a potential conflict point. The observers were asked to categorize each interaction into three risk categories: low (0), medium (1) and high (2).
After the completion of the risk level ratings, Cronbach’s Alpha method was used to evaluate the consistency between all of the observers. This method is a simpler way to determine how consistent the responses are by providing a coefficient (α) of consistency [
41]. The range of the Cronbach’s Alpha coefficient (α) is between 0 and 1, where a higher number denotes a higher level of consistency between observers.
The (α) value in the current study was determined to be 0.82, indicating an acceptable level of consistency [
42,
43].
The entire interaction is divided into three phases in this study; namely, the preliminary phase, peak phase and end phase. The indicators were organized based on a particular phase within the situation development process to which the indicators were most related (refer to
Figure 6 and
Table 4). It should be noted that some indicators are continuous, meaning that they generate values at many points in time and may, therefore, be present at several stages.
The preliminary phase refers to the start of the interaction, which can be understood in several different ways, such as when road users become visible in a camera view or when they begin interacting without any barriers. However, in the context of surrogate safety, the beginning of a conflict is frequently identified as the moment when a road user initiates evasive action, signifying a change from being uninformed of risk to taking measures to prevent it [
20]. Therefore, in this study, the term “preliminary phase” refers to the moment when either of the road users took the first evasive action.
The term “peak phase” refers to the time when the relevant measure reaches its maximum or minimum value in an interaction. As different indicators may reach their maximum or minimum values at different time intervals, multiple values of the peak phase are provided in this study, as can be seen in
Table 4. Lastly, the “end phase” of the interaction refers to the final stage, which is typically the last moment when a collision could still occur, even if theoretically. In this study, this stage refers to the moment when the first road user reaches the potential conflict point.
Model development: A model focused on human perception was created to assess the relationship between the human judgement of severity and surrogate measures. Five trained observers were asked to rank the perceived danger levels of interactions, classifying them as either safe, mild or severe. These ratings were collected using a Likert scale, which is an ordinal scale, meaning that the response variable had more than two categories and these categories had a predefined order. Therefore, an ordinal logistic regression model was used to develop this model.
Equation (2) of ordinal logistic regression can be expressed as below [
44]:
where:
y = response variable;
, ,… = explanatory variables;
c = number of categories for response variable;
cumulative probabilities for category.
P (y > j) = probability that y is greater than j;
= intercepts parameter;
= parameters related to explanatory variable.
This study used the maximum likelihood approach to fit all of the logistic regression models. This test uses a chi-square distribution with n-k-1 degrees of freedom, where n represents the sample size and k represents the number of explanatory variables. The ordinal logistic regression technique was used to develop the above models. The regression coefficients were tested for significance at a level of 95%, which means that any variable with a
p-value of less than 0.05 is considered to be statistically significant. Below,
Table 5,
Table 6,
Table 7 and
Table 8 show the regression co-efficient and goodness of fit test results for models A, B, C and D, which were developed using a forward selection method and only included statistically significant variables at 95% confidence level.
In the first model (Model A), a dataset including all 1141 interactions (with or without evasive action) was used as it was not feasible to estimate indicators for the moment of evasive action for the complete dataset.
Therefore, to determine a value for these indicators for every interaction, the present analysis was constructed with the following goals:
At first, a logistic regression model (model A) was developed using the entire dataset (N = 1141), taking the human judgement of severity as a dependent variable and objective measures that could be estimated for every interaction in the dataset (excluding the indicators estimated at the moment of evasive action) as an independent variable. The aim was to determine if the presence of evasive actions held statistical significance in the model. The model can be represented as below:
MODEL A: Severity rating by human observers (Entire dataset = 1141)~Indicators from moments: (), (), (), (), class variable (EA).
The proxy indicator linked to evasive behavior, as demonstrated in Model A, was deemed significant and ranked second in significance, with the first being the distance. It appears that the pertinent information contained in the earliest conditions of a problem can be used to determine how serious it is.
- 2.
The dataset was then divided into two subgroups: the first subset is where evasive action was observed in one or both road users during an interaction, and the second subset is where no evasive action was observed by either of the road users. Using these subsets of data, two more models (models B and C) were then developed using the same set of explanatory variables. The aim was to examine whether the same indicators were statistically associated with the human judgement of severity in both subsets. The models can be represented as below:
MODEL B: Severity rating by human observers (With Evasive action, N = 905)~Indicators from moments: (), (), (), ().
MODEL C: Severity rating by human observers (With No Evasive action, N = 236)~Indicators from moments: (), (), (), ().
Models B and C were compared with the aim of determining whether the indicators associated with the evasive and non-evasive subsets were similar. The findings, presented in
Table 6 and
Table 7, revealed that
(
) had the most considerable influence on both models, irrespective of the presence of other significant variables at different time intervals during the event, apart from (
).
Thus, it can be inferred that this group of variables could be employed for all events, irrespective of whether there was any evasive action taken or not.
Finally, the last model was developed using only those events where evasive action was observed, and all of the variables presented in
Table 1 were considered. This model aimed to determine whether or not the indicators from the moment of evasive action were helpful in estimating the severity with respect to the human observer ratings. The model can be described as below:
MODEL D: Severity rating by human observers (With Evasive action, N = 905)~Indicators from moments: EA, (), (), (), ().
Models B and D have been compared to evaluate the impact of the variables estimated at the moment of the first evasive action.
Table 6 and
Table 8 reveal that the significant variable sets for both models varied in terms of their emphasis on different stages of the event.
The highest contribution of variables in Model B are the moment of () and PET (), whereas, in Model D, the variables associated with the moment of evasive action (EA) have the greatest importance.
Akaike information criteria and Bayesian information criteria are used in this study to compare the models as the smaller the values, the better the model [
45].
The AIC and BIC scores are defined as below:
where:
From the data in
Table 3, it is evident that Model D is the preferred model when compared to Model B as it had lower AIC and BIC values. Lastly, to validate the models, confusion matrixes were made for all four of the developed models. As can be seen in
Figure 7, given below, the accuracy of all the models was above (95%), with Model D being the highest (98.1%).