Do Traffic Countermeasures Improve the Safety of Vulnerable Road Users at Signalized Intersections? A Combination of Case-Control and Cross-Sectional Studies Using Video-Based Traffic Conflicts

Abstract

Driven by the vision of eliminating road fatalities, Vision Zero initiatives have been widely adopted by many cities around the world, with significant investment of resources in various safety countermeasures. However, there is still a lack of reliable quantitative evidence on the effectiveness of those countermeasures on the traffic conflict frequency at intersections. This research attempts to address this challenge with a combination of case-control and cross-sectional studies, aiming at quantifying the safety effects of three commonly applied Vision Zero countermeasures, namely, leading pedestrian interval, no right turn on red, and installation of a dedicated bicycle lane. A case study was conducted using video trajectory data from 10 signalized intersections in the City of Toronto, Canada. The traffic interactions between vehicles and vulnerable road users were extracted using a video data processing platform, and two surrogate measures of safety, including post-encroachment time and conflict speed, were obtained, and then used to classify the conflict severity into different levels. A comparative analysis using mixed-effects negative binomial regression was conducted to quantify the impacts of different treatments on the frequency of traffic conflicts under specific road weather and traffic conditions. The results show that these three types of traffic countermeasures can effectively reduce the frequency of high-risk and moderate-risk traffic conflicts, moderated by various traffic exposure and weather and environmental conditions and accessible pedestrian signals. These findings could help road safety engineers and decision makers make better informed decisions on their road safety initiatives and projects.

Keywords

safety countermeasure selection quantitative analysis safety project and program evaluation

The Global Status Report on Road Safety 2018 by the World Health Organization (WHO) states that more than half of all road traffic deaths are among vulnerable road users (VRUs), including pedestrians, cyclists, and motorcyclists ( 1 ). Statistics from the city of Toronto, Canada, in 2021, show that pedestrians and cyclists account for 47% of traffic fatalities ( 2 ). Because intersections are the most common sites of traffic conflict between vehicles and VRUs, researchers and traffic managers often focus on intersections for countermeasure implementation ( 3 ). According to the statistical analysis, approximately 28.2% of traffic fatalities take place at intersections ( 4 ).

To address this challenge, Toronto has recently adopted Vision Zero Road Safety Plan—a comprehensive 5-year (2017–2021) action plan focused on reducing traffic-related fatalities and serious injuries on Toronto’s streets. As part of the Plan, several potential traffic countermeasures at intersections have been proposed, focusing on VRUs such as pedestrians and cyclists. Significant resources will be invested to implement these safety improvement projects along with various safety initiatives. There is, therefore, a critical need to develop a data-driven, evidence-based approach to road safety study for assessing the safety benefits of the implemented traffic safety countermeasures at intersections ( 5 , 6 ).

The primary objective of this study is to evaluate the effectiveness of three intersection traffic safety countermeasures: leading pedestrian interval (LPI), no right turn on red (NRTOR), and installation of dedicated bicycle lane (BL). The research applies a conflict-based approach using road user trajectory data, aimed at addressing the following specific questions:

What is the safety effect of a specific safety countermeasure (i.e., LPI, NRTOR, or BL) on the frequency of traffic conflicts with different severity levels?

How do various influencing factors affect the risk of interaction between vehicles and VRUs at signalized intersections?

The rest of this paper is organized as follows. The next section presents a review of related studies, followed by a description of the proposed methodology, including site selection, video data collection and processing, traffic safety indicators selection, and a mixed-effects negative binomial regression model for safety performance evaluation. The section after that provides the results and discussions of the data statistical analysis and modeling. Finally, the contributions of this study to the field of traffic safety evaluation are summarized, the limitations of this study are discussed, and future work is recommended.

Literature Review

Data for Effectiveness Evaluation of Traffic Countermeasures

Depending on the data being used, there are two common methods for evaluating the effectiveness of safety countermeasures: crash-based analysis and surrogate measure of safety (SMoS)-based analysis ( 7 – 13 ).

Crash-based safety study measures the safety level and the effect of a safety countermeasure on accident frequencies by severity ( 14 ). This approach has been widely applied by safety researchers and practitioners for quantifying the safety effects of a variety of safety countermeasures. However, because of the rarity of road crashes, there may not be sufficient crash data available for providing a reliable estimate of the effectiveness of a given treatment within a short period after its implementation. This means that a crash-based analysis could lead to inconclusive or unreliable results because of small sample size ( 15 ). Furthermore, an accident-based safety analysis is reactive in nature, requiring a long period of observation before any action can be taken, which is problematic under the Vision Zero Plan which aims to achieve the state of having no traffic crashes.

The SMoS-based analysis is an alternative method that utilizes observable non-crash events and related attributes, instead of observed crashes, to determine the safety level of a highway facility (e.g., intersection) and the effect of a safety countermeasure ( 10 ). In this approach, surrogate safety measures such as time to collision (TTC) and post-encroachment time (PET) between road users are extracted from user trajectory data captured through simulation or video data collection technology ( 12 ). A certain threshold value is used to identify conflicts or the near miss cases. By comparing these conflicts or near misses before and after the treatment implementation, safety effects can be estimated. The other advantage is that this new approach can continuously collect and analyze video data and generate surrogate safety measures such as speeds, number of interactions, and dangerous conflicts. For instance, Navarro et al. estimated the safety effects of stop signs at intersections using video-based trajectory data and the PET approach ( 12 ). Zangenehpour et al. applied PET extracted from video data to investigate the safety effects of cycle tracks at intersections ( 13 ). Lee et al. evaluated the countermeasures for red light running using a VISSIM simulation model and SMoS which combined TTC and PET ( 11 ). Despite the inherent advantages, very few cities have implemented proactive surrogate safety systems to continuously monitor intersection traffic safety leveraging existing video sensors and existing infrastructure.

Methods for Effectiveness Evaluation of Traffic Countermeasures

The effectiveness evaluation methods could also be classified into the following three types on the basis of the underlying methodology being employed: before-after studies, case-control studies, and cross-sectional studies ( 13 , 16 –21).

A before-after analysis measures outcomes in a group before introducing an intervention, and then again afterwards. For example, Shahdah et al. evaluated the safety performance with different traffic signal operation at intersections through a before-after study based on a traffic microsimulation method ( 17 ). Reyad et al. conducted a before-after study to estimate the treatment effectiveness of traffic signal improvement at intersections using a traffic conflict technique and empirical Bayes model ( 18 ). However, one limitation of before-after studies is that long-period data need to be collected to mitigate the regression-to-the-mean bias. Furthermore, it is difficult to guarantee that the proposed countermeasures are the only changes that occurred at the study site during the period before and after.

A case-control study is often used to identify key factors by comparing two groups (cases and controls). The intersection samples selected for case-control studies typically consist of two subsets: a sample of intersections with countermeasures and a sample of intersections with similar characteristics (e.g., road geometry, traffic volume) but without countermeasures ( 13 , 19 ). Case-control studies can achieve unbiased treatment effects by matching cases and controls for all potential confounding variables.

In cross-sectional studies, count regression models were developed using the data collected from many individual sites at one specific point in time to establish the relationship between traffic crashes or traffic conflicts with variables such as road geometry characteristics, traffic conditions, and implemented safety countermeasures. These models can directly be applied to determine the association of outcomes (e.g., crash, traffic conflicts) with implemented countermeasures without requiring before-after time series data to evaluate the effectiveness of countermeasures ( 20 , 21 ).

However, the weakness of cross-sectional studies is the inability to make causal inferences, and the weakness of case-control studies is the possibility of recall bias. A combination of case-control and cross-sectional approaches provides the benefit of including more sites in the study, making it easier to assess the effectiveness of potential countermeasures at intersections.

Traffic Safety Countermeasures Implemented at Intersections

Several traffic countermeasures have been implemented to improve road safety performance, such as LPI, NRTOR, and installation of dedicated BLs ( 22 – 24 ).

LPI is a safety countermeasure for signalized intersections. With LPI, pedestrians can begin crossing the street before vehicles are given the green light, improving their visibility to the drivers ( 22 ). This is especially needed for pedestrians who are in the blind spot of an approaching vehicle while waiting on the sidewalk curb. By increasing the visibility of pedestrians to drivers, LPI may help reduce the conflicts between turning vehicles and VRUs. Furthermore, this safety countermeasure is also cost-effective to implement, as it only requires a reprogramming of the traffic signals. As shown in Goughnour et al., the LPI treatment can significantly reduce vehicle-pedestrian crashes with an estimated crash modification factor (CMF) of 0.87, where the CMF is a multiplicative factor that indicates the proportion of crashes that would be expected after implementing a countermeasure ( 25 ). NRTOR is a safety intervention aimed at preventing vehicles from making a right turn during a red light ( 23 ). Allowing vehicles to turn right on a red light can have an impact on the safety of pedestrians crossing the street, especially at locations with high pedestrian traffic and areas near schools, parks, and senior homes. In areas with high pedestrian traffic, turning right on red may not be possible because of the constant flow of pedestrians crossing the street, essentially rendering right turns on red useless. One related safety issue that arises is that drivers may fail to stop before the crosswalk and look both ways before slowly creeping up to the intersection to make a right turn on a red. A previous study has also shown that the installation of NRTOR sign can significantly reduce the frequency and severity of traffic conflicts ( 26 ). Lastly, installing dedicated BLs can regulate the paths of cyclists, thereby reducing the traffic conflicts between cyclists and vehicles ( 24 ). As shown in Turner et al., the installation of BLs at signalized intersections can significantly reduce vehicle-bicycle crashes with an estimated CMF of 0.8 ( 27 ).

Methodology

The proposed methodology for evaluating the effectiveness of traffic countermeasures mainly consists of the following four steps: 1) the selection of treatment and control sites; 2) data collection and processing; 3) the identification of traffic conflicts; and 4) the calibration of conflict frequency models and evaluation of the safety effect of countermeasures. A detailed description of these individual steps is provided as follows.

Selection of Treatment and Control Sites

To quantify the safety effect of a countermeasure, this research applies a combination of case-control and cross-sectional studies. Specifically, treatment sites (sites with a countermeasure of interest) and control sites (without any countermeasures) are first identified and safety data from these sites are then collected. This study was motivated by a research project aiming at quantifying the safety effect of several specific safety countermeasures widely implemented in the City of Toronto. These countermeasures were implemented as part of Toronto’s Vision Zero Road Safety Plan—a comprehensive 5-year (2017–2021) action plan focused on reducing traffic-related fatalities and serious injuries on Toronto’s streets. Specifically, three traffic countermeasures are considered in this study: LPI, NRTOR, and installation of dedicated BLs.

Figure 1 shows the 10 selected signalized intersections in city of Toronto with the three different types of traffic countermeasures being implemented: three intersections with NRTOR group, two intersections with LPI, one intersection with a BL on the shoulder of the roadway, and four intersections as part of the control group. All intersections were chosen in consultation with City of Toronto staff with the following two considerations: 1) the selected locations have the safety countermeasures of interest and 2) the selected intersections have similar geometric characteristics and traffic and environmental conditions. The intersections selected in this study are all signal intersections with four legs, the same posted speed limit of 50 km/h, and five lanes in each leg, as shown in Figure 1.

Figure 1.

Selected intersections with implemented traffic countermeasures.

Data Collection and Processing

A total of 1,118 h of video data at the 10 selected intersections were collected: 331, 237, 118, and 432 h for the NRTOR, LPI, BL and control groups, respectively. All video data were then processed using a commercial software TrafxSAFE provided by Transoft to extract road user trajectories and conflict data to perform conflict analyses.

The video data processing procedure has four steps: 1) detecting and tracking the road users in the video; 2) classifying the detected road users into different types, that is, vehicle, pedestrian, and cyclist; 3) selecting the interactive road user pairs with traffic conflict; and 4) calculating the surrogate measures of safety for extracted traffic conflict ( 28 ).

The software records data for each conflict, including conflicting traffic (e.g., vehicle type and pedestrian), PET, and speeds. To prevent the cameras from missing some of the conflicts that occurred, the locations of the cameras must meet certain requirements, such as video resolution, field of view, and height/distance from the intersection. All traffic conflicts with PET less than 10 s were recorded for this study. Figure 2 shows the conflicts by severity for a conflicting movement and a scenario obtained from the data analysis platform. Red and blue in the heat map indicate the frequency of the detected traffic conflicts, where red indicates a higher number of conflicts.

Figure 2.

Outcomes from the video analytics platform.

A total of 101,364 traffic conflicts (i.e., PET <10 s) at 10 selected intersections were recorded and used in this study. Table 1 shows the summary statistics of the collected variables for the evaluation of the effectiveness of traffic countermeasures. Since the selected intersections are all signalized, this paper mainly focuses on the traffic conflicts between right-turning vehicles and pedestrians or cyclists on the crosswalk. The conflict variables include conflict speed and the PET value for that conflict. Road user 1 represents the conflicting vehicle, which includes cars, work vans, buses, motorcycles, and so forth, while road user 2 is a pedestrian or a bicycle in the crosswalk or BL. Besides, hourly right-turn vehicle flow and hourly crosswalk flow are included in the traffic volume variables. Among them, the crosswalk volume includes the volume of pedestrians and bicycles on the crosswalk or BL. In addition, climate data including temperature, humidity, wind speed, visibility, and weather condition, were retrieved from Environment Canada’s historical weather database, using the Toronto City Centre Climate Station. The climate condition at each hour was associated with each observation—this is used to determine if poor weather condition was a factor affecting the occurrence and severity of traffic conflicts. Furthermore, accessible pedestrian signals (APS) information for each selected intersection were also collected and included in our modeling.

Table 1.

Summary of Variables Descriptive Statistics

Variables	Mean	SD	Min.	Max.
Conflict variables
Road user 1 conflict speed (km/h)	18.297	8.675	0	59.990
Road user 2 conflict speed (km/h)	7.070	3.911	0	16.790
PET (s)	5.223	1.949	0	9.980
Traffic volume variables
Right turn volume per hour	295.932	177.039	9	874
Crosswalk volume per hour	478.579	452.306	2	2109
Climate variables
Temperature (°C)	−1.658	3.654	−10.300	8.800
Humidity (%)	72.616	10.982	39	93
Wind speed (km/h)	20.847	9.384	0	46
Visibility (km)	13.960	4.426	1.600	16.100
Variables	Dummy variable		Number of traffic conflicts
Weather condition	Sunny		71,293
	Snow		25,034
	Rain		2,333
	Fog		556
	Haze		2,148
Temporal variables
Time-of-day	Daytime		81,253
Time-of-day	Nighttime		20,111
Traffic signal variables
APS	No APS signal		25,553
	Pushbutton actuated		56,119
	Fixed time		19,692
Countermeasures variables
Countermeasure	Control		39,697
	LPI		20,698
	NRTOR		30,075
	BL		10,894

Note: APS = accessible pedestrian signals; BL = bicycle lane; LPI = leading pedestrian interval; Max. = maximum; Min. = minimum; NRTOR = no right turn on red; PET = post-encroachment time; SD = standard deviation.

Identification of Traffic Conflicts of Varying Levels of Severity

In this study, traffic conflicts are defined on the basis of two safety surrogate measures jointly: road user conflict speed and PET, following the methodology by Navarro et al. and Zangenehpour et al. ( 12 , 13 ). The threshold values for road user conflict speed were inspired by the statistical results conducted by the European Traffic Safety Council (ETSC), which states that the probability of pedestrian fatality is 5%, 45%, and 85%, respectively, when hit by a vehicle at 32 km/h, 48 km/h, and 64 km/h ( 29 ). Zangenehpour et al. classified conflicts into four categories based on PET ( 13 ):

Very dangerous interaction: PET ≤ 1.5 s

Dangerous interaction: 1.5 s < PET ≤ 3 s

Mild interaction: 3 s < PET ≤ 5 s

No interaction: PET > 5 s

As a result, four severity levels are defined for interactions or conflicts between vehicles and VRUs based on conflict speed and PET, as shown in Figure 3. Four categories are shown as below ( 12 ):

High-risk interaction: PET ≤ 1.5 s and vehicle speed > 48 km/h

Moderate-risk interaction: PET ≤ 1.5 s and 32 km/h < vehicle speed ≤ 48 km/h, or 1.5 s < PET ≤ 3 s and vehicle speed > 32 km/h

Low-risk interaction: PET ≤ 3 s and 16 km/h < vehicle speed ≤ 32 km/h, or 3 s < PET ≤ 5 s and vehicle speed > 16 km/h

Safe interaction: PET ≤ 5 s and 0 km/h < vehicle speed ≤ 16 km/h, or 5 s < PET ≤ 10 s and vehicle speed > 0 km/h

In this study, two traffic conflict severity classification methods (i.e., PET-based, PET and conflict speed-based) were also compared in the traffic countermeasures evaluation.

Figure 3.

Classification of severity levels for vehicle and vulnerable road user interactions ( 12 ).

In addition, the traffic conflict rate considering the conflict frequency and traffic volume is calculated using the following equation 1 ( 13 ). The Kruskal-Wallis test, a non-parametric analysis method which has the advantage that the samples do not need to satisfy a specific distribution, was used to test the difference in traffic conflict rates between different countermeasure groups ( 30 ).

C R_{i} = \frac{C F_{i} \times 10^{6}}{R i g h t T u r n V o l u m e P e r H o u r \times C r o s s w a l k V o l u m e P e r H o u r}

(1)

where

$C R_{i}$ = the conflict rate of severity level $i$ , and

$C F_{i}$ = hourly traffic conflict frequency of severity level $i$ .

Calibration of Conflict Frequency Models and Evaluation of Countermeasure Effects

In this study, the hourly traffic conflict frequency of different severity levels was chosen as the measure of safety level of the selected intersections. As discussed in the literature review, various statistical methods have been applied to model traffic conflict frequency over the past several years, such as the Poisson regression model and the negative binomial regression model ( 17 , 31 ). However, traditional Poisson and negative binomial regression models assume that the traffic conflicts at intersections are independent, thus ignoring the heterogeneity characteristics of different intersections. In this research, a mixed-effects negative binomial regression model is introduced, which is known to be effective in modeling varied effects between intersections with repeated measurements ( 32 ).

The mixed-effects negative binomial model was adopted by introducing a random intersection-specific effects term into the relationship between the expected number of traffic conflicts at a given time period $t$ at a given intersection i, ${\tilde{μ}}_{it}$ ,and its explanatory variables $X_{it}$ , as shown below:

{\tilde{μ}}_{it} = μ_{it} δ_{i}

(2)

where

$δ_{i}$ = the random location-specific effects term.

Similar to the negative binomial model, the value of ${\tilde{μ}}_{it}$ is guaranteed positive values by assuming:

{\tilde{μ}}_{it} = μ_{it} δ_{i} = \exp (β X_{it} + μ_{i})

(3)

where

$β$ = the estimated coefficient vector of the applied model,

$\exp (μ_{i})$ = a gamma-distributed term with mean 1 and variance $k$ (where $k$ is the overdispersion parameter in the negative binomial model).

Therefore, the probability density function of the mixed-effects negative binomial regression model can be represented as:

f (y_{it} | X_{it}) = \frac{Γ (a + b) Γ (a + \sum_{T} μ_{it}) Γ (b + \sum_{T} μ_{it})}{Γ (a) Γ (b) Γ (a + b + \sum_{T} μ_{it} + \sum_{T} y_{it})} \underset{T}{Π} \frac{Γ (μ_{it} + y_{it})}{Γ (μ_{it}) + Γ (y_{it} + 1)}

(4)

where

$y_{it}$ = the number of traffic conflicts at intersection $i$ during given period $t$ , and

$a$ and $b$ = distributional parameters (which can be estimated using the maximum likelihood method).

The likelihood ratio (LR) test was applied in this study to compare the performance of the mix-effects negative binomial model and the traditional fixed negative binomial model. The incidence-rate ratio (IRR) for each explanatory variable was also calculated to represent the effect of changes in that variable on the model results. All coefficients in the proposed mixed-effects negative binomial model were estimated using Stata/MP 16.0.

Results and Discussion

Comparative Analysis of Traffic Conflict Frequency under Different Countermeasures

An initial investigation of the conflict data collected by the video equipment was conducted and the PET values of individual traffic conflicts for each treatment group were analyzed. Table 2 shows the summary description of the processed video data for the selected intersections with different countermeasures. The average traffic conflict rates at different severity levels are counted and provided in Table 2 and Figure 4. Table 3 provides the results of the Kruskal-Wallis test of average traffic conflict rates of different risk levels under different traffic countermeasures.

Table 2.

Summary Description of the Processed Video Data for Intersections with Implemented Traffic Countermeasures

		Average traffic conflict rate						Turn speed (km/h)		Cross speed (km/h)		Volume (per hour)
Intersection		High-risk	Moderate-risk	Low-risk	Very dangerous PET ≤ 1.5 s	Dangerous 1.5 s < PET ≤ 3 s	Mild 3 s < PET ≤ 5 s	Mean	SD	Mean	SD	Turn	Cross
Control	Sheppard/Victoria Park	0.14	11.40	256.50	2.12	110.02	268.22	22.70	10.01	7.76	4.55	460.43	289.45
	Yonge/Finch	0.23	4.13	224.51	2.73	80.44	356.17	18.23	8.82	7.22	4.27	226.56	751.18
	Danforth/Pape	0.02	2.92	191.79	1.45	73.23	525.81	15.75	8.88	6.28	3.15	201.32	655.38
	Eglinton/Martin Grove	0	3.26	247.89	6.97	71.93	272.9	19.80	6.48	7.28	5.14	320.17	60.82
LPI	Lawrence/Don Mills	0	12.95	282.31	2.62	79.68	333.22	22.21	9.07	7.08	4.19	426.09	255.44
LPI	Bloor/Jane	0	1.35	392.15	3.36	109.37	688.36	16.48	6.46	6.40	2.54	385.95	243.89
NRTOR	Spadina/Dundas	0.07	1.36	214.52	1.16	171.96	503.75	15.10	7.02	7.11	3.77	145.83	838.44
	Bloor/Parkside	0	3.44	896.48	3.14	138.23	1153.20	19.26	6.79	6.00	3.17	239.60	177.44
	Bloor/Avenue	0.05	30.92	219.62	4.21	162.31	260.27	18.83	9.78	8.00	4.40	286.09	743.66
BL	King/Strachan	0	1.45	529.12	3.00	152.56	973.04	16.57	5.85	7.23	3.63	138.28	385.44
Total	Control	0.10	5.52	229.73	3.20	84.41	357.98	18.79	9.49	7.05	4.12	302.12	439.21
	LPI	0	6.38	344.54	3.04	96.50	534.42	18.91	8.18	6.69	3.36	403.57	248.96
	NRTOR	0.04	11.23	413.09	2.69	159.19	614.95	17.85	8.62	7.30	4.05	219.77	626.26
	BL	0	1.45	529.12	3.00	152.56	973.04	16.57	5.85	7.23	3.63	138.28	385.44

Note: BL = bicycle lane; LPI = leading pedestrian interval; NRTOR = no right turn on red; PET = post-encroachment time; SD = standard deviation.

Figure 4.

Box-violin plot of traffic conflict rate with different severity levels under different traffic countermeasures.

Table 3.

Results of Kruskal-Wallis Test of Traffic Conflict Rate with Different Risk Levels under Different Traffic Countermeasures

Group 1	Group 2	High-risk interaction		Moderate-risk interaction		Low-risk interaction
Group 1	Group 2	$χ^{2}$	p	$χ^{2}$	p	$χ^{2}$	p
Control	LPI	6.94	0.008***	5.46	0.019**	25.58	<0.001***
Control	NRTOR	1.55	0.213	5.96	0.015**	0.07	0.791
Control	BL	3.38	0.066*	28.89	<0.001***	130.56	<0.001***
LPI	NRTOR	3.64	0.056*	0.03	0.867	14.38	<0.001***
LPI	BL	na	na	13.52	<0.001***	81.90	<0.001***
NRTOR	BL	1.77	0.183	16.82	<0.001***	57.25	<0.001***
Total		10.40	0.016**	31.63	<0.001***	130.53	<0.001***
Group 1	Group 2	Very dangerous interaction		Dangerous interaction		Mild interaction
Group 1	Group 2	$χ^{2}$	p	$χ^{2}$	p	$χ^{2}$	p
Control	LPI	3.72	0.054*	0.15	0.702	16.86	<0.001***
Control	NRTOR	0.62	0.431	31.61	<0.001***	0.78	0.376
Control	BL	0.37	0.544	111.14	<0.001***	147.57	<0.001***
LPI	NRTOR	1.75	0.185	27.20	<0.001***	3.05	0.081*
LPI	BL	0.66	0.416	93.17	<0.001***	79.94	<0.001***
NRTOR	BL	0.09	0.762	62.32	<0.001***	57.32	<0.001***
Total		3.92	0.270	144.05	<0.001***	129.36	<0.001***

Note: BL = bicycle lane; LPI = leading pedestrian interval; NRTOR = no right turn on red; na = not applicable.

significant at 10% level

significant at 5% level

***

significant at 1% level

As provided in Table 2, the turning movement speed was lower in the NRTOR and BL groups than in the control group, suggesting that the countermeasures generally reduced the speed of turning movements, and that they have therefore contributed to the reduction in the severity of the potential traffic conflicts. As shown in Figure 4 and Table 3, the high-risk interaction rate at the intersections treated with LPI and BL was significantly lower than that in the control group. As shown in Table 2 and Figure 4, the average high-risk conflict rate reduced from 0.1 in the control group to 0 in the LPI and BL groups. The average moderate-risk conflict rate reduced from 5.52 in the control group to 1.45 in the BL group. Furthermore, the average low-risk interaction rate at the intersections treated with LPI and BL was significantly higher than that in the control group. Table 3 shows the very dangerous interaction rate was significantly lower in the LPI group than in the control group when traffic conflict levels were classified using PET values. As presented in Table 2 and Figure 4, the average dangerous interaction rate was reduced from 3.2 in the control group to 3.04 in the LPI group. At the same time, the dangerous interaction rate in the NRTOR group and the BL group was significantly higher than that in the control group.

Safety Effects of Countermeasures

To evaluate the effectiveness of the three countermeasures (i.e., LPI, NRTOR, and BL), separate conflict frequency models were developed using mixed-effects negative binomial regression. Tables 4 and 5 present the modeling results for conflict frequencies of different severity classes defined by the two approaches introduced previously: PET-based and PET-with-conflict-speed-based. The mixed-effects model performed better than the fixed-effects model in most comparison groups. All insignificant variables were eliminated in the final model.

Table 4.

Results of Traffic Conflict (Classified by PET Value and Conflict Speed) Frequency Modeling

Type of countermeasure	Control-LPI		Control-NRTOR		Control-BL
Explanatory variables	$β$	IRR	$β$	IRR	$β$	IRR
High-risk interaction frequency modeling
Fixed effects
Constant	−9.256***	0.000	−7.256***	0.001	−9.270***	0.000
Right turn volume per hour	0.006**	1.006	0.006***	1.006	0.008***	1.008
Crosswalk volume per hour	0.004***	1.004	0.003***	1.003	0.004***	1.004
Countermeasure (baseline: control)	na	na	−1.796**	0.166	na	na
Random effects (group name: intersection)
Variance (constant)	0.765		na		na
Log likelihood	−46.856		−68.467		−44.050
LR test versus negative binomial model	$χ^{2}$ =1.83, P=0.08		na		na
Moderate-risk interaction frequency modeling
Fixed effects
Constant	−2.454***	0.086	−1.348*	0.260	−2.859***	0.057
Right turn volume per hour	0.004***	1.004	0.003***	1.003	0.003***	1.003
Crosswalk volume per hour	0.002***	1.002	0.001***	1.001	0.002***	1.002
Time-of-day—nighttime (baseline: daytime)	−0.557***	0.573	−0.486***	0.615	−0.690***	0.502
Humidity (%)	na	na	0.011**	1.011	0.010*	1.010
AP—pushbutton actuated (baseline: no APS)	na	na	−1.214*	0.297	na	na
APS—fixed time (baseline: no APS)	na	na	−1.201*	0.301	na	na
Countermeasure (baseline: control)	−0.611***	0.543	−0.731**	0.481	−0.931*	0.394
Random effects (group name: intersection)
Variance (constant)	0.413		0.271		0.338
Log likelihood	−583.031		−702.091		−443.303
LR test versus negative binomial model	$χ^{2}$ =76.99, P < 0.01		$χ^{2}$ =48.14, P <. 0.01		$χ^{2}$ =47.02, P < 0.01
Low-risk interaction frequency modeling
Fixed effects
Constant	1.951***	7.033	1.597***	4.937	2.677***	14.535
Right turn volume per hour	0.003***	1.003	0.004***	1.004	0.004***	1.004
Crosswalk volume per hour	0.001***	1.001	0.001***	1.001	0.001***	1.001
Time-of-day—nighttime (baseline: daytime)	−0.578***	0.561	−0.658***	0.518	−0.477***	0.621
Humidity (%)	−0.008***	0.992	−0.005***	0.995	−0.006***	0.994
Wind speed (km/h)	−0.004**	0.996	na	na	−0.003*	0.997
Visibility (km)	−0.006**	0.994	na	na	na	na
APS—pushbutton actuated (baseline: no APS)	na	na	na	na	−1.145***	0.318
APS—fixed time (baseline: no APS)	na	na	na	na	−0.684*	0.504
Countermeasure (baseline: control)	0.362***	1.436	0.444**	1.559	na	na
Random effects (group name: intersection)
Variance (constant)	0.074		0.245		0.081
Log likelihood	−2,158.765		−2,520.843		−1,729.532
LR test versus negative binomial model	$χ^{2}$ =152.44, P < 0.01		$χ^{2}$ =375.86, P <. 0.01		$χ^{2}$ =125.51, P < 0.01

Note: APS = accessible pedestrian signals; BL = bicycle lane; IRR = incidence-rate ratio; LPI = leading pedestrian interval; LR = likelihood ratio; NRTOR = no right turn on red; PET = post-encroachment time; na = not applicable.

significant at 10% level

significant at 5% level

***

significant at 1% level

Table 5.

Results of Traffic Conflict (Classified by PET Value) Frequency Modeling

Type of countermeasure	Control-LPI		Control-NRTOR		Control-BL
Explanatory variables	$β$	IRR	$β$	IRR	$β$	IRR
Very dangerous interaction frequency modeling
Fixed effects
Constant	−3.577***	0.028	−3.501***	0.030	−3.791***	0.023
Right turn volume per hour	0.004**	1.003	0.005***	1.005	0.005***	1.004
Crosswalk volume per hour	0.002***	1.002	0.001***	1.001	0.002***	1.002
Weather—rain (baseline: snow)	1.813***	6.133	na	na	na	na
Countermeasure (baseline: control)	−0.466**	0.627	na	na	−0.752***	0.471
Random effects (group name: intersection)
Variance (constant)	na		0.065		na
Log likelihood	−403.679		−523.400		−347.003
LR test versus negative binomial model	na		$χ^{2}$ =4.22, P=0.02		na
Dangerous Interaction Frequency Modeling
Fixed Effects
Constant	0.173*	1.189	na	na	na	na
Right turn volume per hour	0.003***	1.003	0.004***	1.004	0.003***	1.003
Crosswalk volume per hour	0.002***	1.002	0.001***	1.001	0.002***	1.002
Time-of-day—nighttime (baseline: daytime)	−0.391***	0.676	−0.439***	0.645	−0.318***	0.727
Countermeasure (baseline: control)	na	na	0.585*	1.796	1.022***	2.779
Random effects (group name: intersection)
Variance (constant)	0.106		0.354		0.154
Log likelihood	−1,519.009		−1,859.096		−1,252.479
LR test versus negative binomial model	$χ^{2}$ =86.26, P < 0.01		$χ^{2}$ =281.19, P < 0.01		$χ^{2}$ =65.83, P < 0.01
Mild interaction frequency modeling
Fixed Effects
Constant	0.107*	1.113	na	na	na	na
Right turn volume per hour	0.003***	1.003	0.004***	1.004	0.003***	1.003
Crosswalk volume per hour	0.002***	1.002	0.001***	1.001	0.002***	1.002
Time-of-day—nighttime (baseline: daytime)	−0.388***	0.678	−0.432***	0.649	−0.307***	0.736
APS—fixed time (baseline: no APS)	0.466***	1.593	na	na	na	na
Countermeasure (baseline: control)	na	na	0.578*	1.782	1.016***	2.763
Random effects (group name: intersection)
Variance (constant)	0.063		0.343		0.148
Log likelihood	−1,544.233		−1,885.253		−1,272.367
LR test versus negative binomial model	$χ^{2}$ =59.81, P < 0.01		$χ^{2}$ =275.49, P < 0.01		$χ^{2}$ =60.97, P < 0.01

Note: APS = accessible pedestrian signals; BL = bicycle lane; LPI = leading pedestrian interval; LR = likelihood ratio; NRTOR = no right turn on red; PET = post-encroachment time; na = not applicable.

significant at 10% level

significant at 5% level

***

significant at 1% level

When PET and conflict speed are combined to classify conflict severity, as shown in Table 4, it can be found that right-turn traffic volume and crosswalk volume have a positive impact on the frequency of traffic conflict at any severity level. Furthermore, the results of high-risk interaction frequency modeling showed that only the NRTOR group significantly reduced the frequency of high-risk interactions between vehicles and VRUs, and the high-risk interaction frequency of the NRTOR group decreased by 83.4% compared with the control group. As for the results of the moderate-risk interaction frequency modeling, it can be found that the frequency of moderate-risk conflicts is significantly reduced at night compared with daytime, the same finding as in the results of low-risk interaction frequency modeling. In addition, compared with the control group, the three traffic countermeasures—LPI, NRTOR, and BL—all significantly reduced the moderate-risk interaction frequency by 45.7%, 51.9%, and 60.6%, respectively. As shown in the control-NRTOR group modeling results, humidity was positively correlated with moderate-risk interaction frequency. Meanwhile, both pushbutton-actuated and fixed-time APS can significantly reduce the moderate-risk interaction frequency compared with no APS. For the results of low-risk interaction frequency modeling, compared with the control group, the frequency of low-risk interaction in the LPI and NRTOR groups significantly increased by 43.6% and 55.9%, respectively. Environmental factors such as humidity, wind speed, and visibility all negatively affect the frequency of low-risk interactions. As shown in the results of the control-BL group modeling, the frequency of low-risk interactions was lower in the APS-equipped intersections than in the no-APS intersections. The variance of intercept increases considerably from the low-risk to high-risk interaction category. The main reason is that, when considering conflict speed and PET for conflict severity classification, the number of high-risk samples is much smaller and the degree of dispersion of each intersection is larger, which may have led to an increase in the variance of the intercept.

Table 5 shows the modeling results of interaction frequency when only the PET value is used for interaction severity classification. Similar to the findings in Table 4, both right-turn traffic volume and crosswalk volume positively affect the frequency of various types of traffic interaction between vehicles and VRUs. Compared with the control group, the LPI and BL groups showed significant reductions in the frequency of very dangerous interactions by 37.3% and 52.9%, respectively. As shown in the results of the control-LPI group, the frequency of very dangerous interactions increased by 513.3% on rainy days compared with sunny days. For the results shown in the dangerous interaction frequency modeling, compared with the control group, the NRTOR and BL groups showed a significant increase in the frequency of dangerous interactions by 79.6% and 177.9%, respectively. In addition, the three comparison modeling results showed that, compared with the daytime, the frequency of dangerous interactions was reduced by 32.4%, 35.5%, and 27.3% at nighttime, respectively. Lastly, as shown in the results of mild interaction frequency modeling, compared with the control group, the frequency of mild interactions at the intersections with NRTOR and BL countermeasures significantly increased by 78.2% and 176.3%, respectively. At the same time, the three comparison modeling results showed that, compared with daytime, the frequency of mild interactions at nighttime was reduced by 32.2%, 35.1%, and 26.4%, respectively. Meanwhile, as shown in the control-LPI group modeling results, intersections with fixed-time APS devices have a higher frequency of mild interactions than intersections without APS.

Conclusion

This paper introduced an evidence-based study combining case-control and cross-sectional approaches to quantify the effectiveness of three widely used intersection safety countermeasures targeting VRUs—LPI, NRTOR, and installation of dedicated BLs. Video-based trajectory data from 10 selected intersections in the City of Toronto were used to determine two surrogate traffic safety indicators including PET and conflict speed. A mixed-effects negative binomial regression model was applied to quantify the impact of different traffic countermeasures on the traffic conflict frequency of the intersections. Three key conclusions can be drawn from our investigation as follows:

Based on both PET-based or conflict-speed-based conflict severity analyses, the frequency of the highest risk interactions between vehicles and VRUs at intersections with the three traffic countermeasures (i.e., LPI, NRTOR, and BL) was consistently lower than those at the intersections without these safety countermeasures.

The mixed-effects negative binomial regression model worked well in quantifying the impact of the traffic safety countermeasures on intersection safety. Compared with the control group, the frequency of high-risk interactions at intersections with NRTOR countermeasures decreased by 83.4%. Meanwhile, the LPI group and the BL group showed significant reductions in the frequency of critically dangerous interactions by 36.3% and 52.9%, respectively. In addition, the three traffic safety countermeasures also reduced the frequency of moderate-risk interactions by 45.7%, 51.9%, and 60.6%, respectively.

Among different explanatory variables, traffic volume plays a key role in influencing the frequency of traffic conflicts at any severity level. Furthermore, time-of-day and some meteorological variables, such as weather, humidity, wind speed, and visibility, also significantly affect the frequency of certain types of traffic conflict. In addition, different types of APS also have a significant impact on the safety performance of intersections.

This study has provided a theoretical basis for conducting a data-driven effectiveness evaluation of specific safety countermeasures. However, several limitations still exist in this study. Because of the lack of a sufficiently long history of traffic crash data at the selected intersections, the traffic safety indicators applied in this paper have not been validated. Because of the limitation of the data collection period, the sample of high-risk interactions of some treatment groups is still limited and the generalizability of the conclusions obtained in this study still needs to be further verified. In future work, real-world crash data should be collected and used to validate the effectiveness of the traffic safety indicators. In addition, safety diagnosis of different types of intersections and different conflict scenarios should also be carried out to form a more general knowledge base on the effectiveness of various traffic safety countermeasures. Furthermore, this study did not distinguish between vehicle-pedestrian conflicts and vehicle-bicycle conflicts, which may have different conflict patterns, and different traffic control measures may have different impacts on them. Future research will further explore the impact of different countermeasures on different conflict types. Lastly, this study mainly focuses on the traffic conflict between right-turning vehicles and pedestrians in crosswalks without considering the possible conflicts between left-turning vehicles and pedestrians. The impact of left-turning traffic volume on the efficiency of LPI treatment is also worthy of further exploration.

Footnotes

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: Q. Shangguan, J. Keung, L. Fu, L. Samara, J. Wang, T. Fu; data collection: Q. Shangguan, J. Keung; analysis and interpretation of results: Q. Shangguan, J. Keung, L. Samara, L. Fu; draft manuscript preparation: Q. Shangguan, J. Keung, L. Fu, L. Samara, J. Wang, T. Fu. All authors reviewed the results and approved the final version of the manuscript.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was jointly supported by a City of Toronto project entitled “A Pilot Project for Continuous Video-Based Traffic Analysis and Safety Monitoring,” Ontario Research Fund – Research Excellence (ORF-RE) project “Intelligent Systems for Sustainable Urban Mobility (ISSUM),” and the program of China Scholarship Council (202106260118). Transoft provided technology and support for video data processing.

ORCID iDs

Qiangqiang Shangguan

Liping Fu

Lana Samara

Ting Fu

Data Accessibility Statement

The data that support the findings of this study are available from the corresponding author, Liping Fu, on reasonable request.

References

WHO. Global Status Report on Road Safety 2018. Geneva, 2018.

City of Toronto. Fatalities – Vision Zero. https://www.toronto.ca/services-payments/streets-parking-transportation/road-safety/vision-zero/vision-zero-dashboard/fatalities-vision-zero/.

Ghoul

Sayed

Real-Time Signal-Vehicle Coupled Control: An Application of Connected Vehicle Data to Improve Intersection Safety. Accident Analysis & Prevention, Vol. 162, 2021, p. 106389.

Federal Highway Administration. Intersection Safety. Washington, D.C., 2022. https://safety.fhwa.dot.gov/intersection/.

Fleisher

Wier

M. L.

Hunter

A Vision for Transportation Safety: Framework for Identifying Best Practice Strategies to Advance Vision Zero. Transportation Research Record: Journal of the Transportation Research Board, 2016. 2582: 72–86.

Kronenberg

Woodward

DuBose

Weissman

Achieving Vision Zero: Data-Driven Investment Strategy to Eliminate Pedestrian Fatalities on a Citywide Level. Transportation Research Record: Journal of the Transportation Research Board, 2015. 2519: 146–156.

Kay

Gates

T. J.

Savolainen

P. T.

Shakir Mahmud

Safety Performance of Unsignalized Median U-Turn Intersections. Transportation Research Record: Journal of the Transportation Research Board, 2022. 2676: 451–466.

Paul

Ghosh

Mazharul Haque

The Effects of Green Signal Countdown Timer and Retiming of Signal Intervals on Dilemma Zone Related Crash Risk at Signalized Intersections Under Heterogeneous Traffic Conditions. Safety Science, Vol. 154, 2022, p. 105862.

Sasidharan

Donnell

E. T.

Application of Propensity Scores and Potential Outcomes to Estimate Effectiveness of Traffic Safety Countermeasures: Exploratory Analysis Using Intersection Lighting Data. Accident Analysis & Prevention, Vol. 50, 2013, pp. 539–553.

10.

El-Basyouny

Sayed

Safety Performance Functions Using Traffic Conflicts. Safety Science, Vol. 51, No. 1, 2013, pp. 160–164.

11.

Lee

Evaluation of Countermeasures for Red Light Running by Traffic Simulator–Based Surrogate Safety Measures. Traffic Injury Prevention, Vol. 19, No. 1, 2018, pp. 1–8.

12.

Navarro

Miranda-Moreno

Saunier

Labbe

Do Stop-Signs Improve the Safety for All Road Users? A Before-After Study of Stop-Controlled Intersections Using Video-Based Trajectories and Surrogate Measures of Safety. Accident Analysis & Prevention, Vol. 167, 2022, p. 106563.

13.

Zangenehpour

Strauss

Miranda-Moreno

L. F.

Saunier

Are Signalized Intersections with Cycle Tracks Safer? A Case–Control Study Based on Automated Surrogate Safety Analysis Using Video Data. Accident Analysis & Prevention, Vol. 86, 2016, pp. 161–172.

14.

National Research Council. Highway Safety Manual. AASHTO, Washington, D.C., 2010.

15.

Van Raemdonck

Macharis

Mairesse

Risk Analysis System for the Transport of Hazardous Materials. Journal of Safety Research, Vol. 45, 2013, pp. 55–63.

16.

Fayish

A. C.

Gross

Safety Effectiveness of Leading Pedestrian Intervals Evaluated by a Before–After Study with Comparison Groups. Transportation Research Record: Journal of the Transportation Research Board, 2010. 2198: 15–22.

17.

Shahdah

Saccomanno

Persaud

Application of Traffic Microsimulation for Evaluating Safety Performance of Urban Signalized Intersections. Transportation Research Part C: Emerging Technologies, Vol. 60, 2015, pp. 96–104.

18.

Reyad

Sacchi

Ibrahim

Sayed

Traffic Conflict–Based Before–After Study with Use of Comparison Groups and the Empirical Bayes Method. Transportation Research Record: Journal of the Transportation Research Board, 2017. 2659: 15–24.

19.

Abdulsalam

A. J.

Rowlands

Easa

S. M.

Abd El Halim

A. E. H. O.

Novel Case-Control Observational Method for Assessing Effectiveness of Red-Light Cameras. Canadian Journal of Civil Engineering, Vol. 44, No. 6, 2017, pp. 407–416.

20.

Zegeer

Lyon

Srinivasan

Persaud

Lan

Smith

Carter

Thirsk

N. J.

Zegeer

Ferguson

Development of Crash Modification Factors for Uncontrolled Pedestrian Crossing Treatments. Transportation Research Record: Journal of the Transportation Research Board, 2017. 2636: 1–8.

21.

Dissanayake

Shams

Safety Evaluation of Shoulder Bypass Lanes at Unsignalized Intersections on Rural Two-Lane Roadways Using Cross Sectional Analysis. Journal of the Transportation Research Forum, Vol. 55, No. 3, 2016. pp. 5–22.

22.

Guo

Sayed

Zheng

A Hierarchical Bayesian Peak Over Threshold Approach for Conflict-Based Before-After Safety Evaluation of Leading Pedestrian Intervals. Accident Analysis & Prevention, Vol. 147, 2020, p. 105772.

23.

Ellis

Van Houten

Reduction of Pedestrian Fatalities, Injuries, Conflicts, and Other Surrogate Measures in Miami–Dade, Florida: Results of Large-Scale FHWA Project. Transportation Research Record: Journal of the Transportation Research Board, 2009. 2140: 55–62.

24.

Goh

K. C. K.

Currie

Sarvi

Logan

Road Safety Benefits from Bus Priority: An Empirical Study. Transportation Research Record: Journal of the Transportation Research Board, 2013. 2352: 41–49.

25.

Goughnour

Carter

Lyon

Persaud

Lan

Chun

Hamilton

Signor

Bryson

Evaluation of Protected Left-Turn Phasing and Leading Pedestrian Intervals Effects on Pedestrian Safety. Transportation Research Record: Journal of the Transportation Research Board, 2021. 2675: 1219–1228.

26.

Sayed

Ismail

Zaki

M. H.

Autey

Feasibility of Computer Vision-Based Safety Evaluations: Case Study of a Signalized Right-Turn Safety Treatment. Transportation Research Record: Journal of the Transportation Research Board, 2012. 2280: 18–27.

27.

Turner

Wood

Hughes

Singh

Safety Performance Functions for Bicycle Crashes in New Zealand and Australia. Transportation Research Record: Journal of the Transportation Research Board, 2011. 2236: 66–73.

28.

Transoft Solutions. https://safety.transoftsolutions.com/trafxsafe/.

29.

European Traffic Safety Council. European Road Safety Observatory (2006) Speeding, European Traffic Safety Council, Brussels, 2007.

30.

McKight

P. E.

Najab

Kruskal-Wallis Test. In: The Corsini Encyclopedia of Psychology ( Weiner

I. B.

Craighead

W. E.

, eds.), John Wiley & Sons Inc, Hoboken, NJ, 2010, pp. 1–1.

31.

Zhang

Chien

S. I. J.

Wang

Dong

Predicting Expressway Crash Frequency Using a Random Effect Negative Binomial Model: A Case Study in China. Accident Analysis & Prevention, Vol. 98, 2017, pp. 214–222.

32.

Stapleton

S. Y.

Gates

T. J.

Avelar

Geedipally

S. R.

Saedi

Safety Performance Functions for Low-Volume Rural Stop-Controlled Intersections. Transportation Research Record: Journal of the Transportation Research Board, 2019. 2673: 660–669.