In an article published in the journal Scientific Reports, researchers from the UK and Australia collaboratively developed an innovative framework for real-time estimation and prediction of crash risk at signalized intersections. Their technique holds the potential to offer novel insights for proactive safety management and the implementation of risk-responsive countermeasures at these intersections.
Study: AI-Driven Framework for Real-Time Crash Risk Forecasting at Intersections. Image credit: Supitcha McAdam/Shutterstock
Background
Signalized intersections are pivotal points for traffic safety due to the complex interactions among vehicles, pedestrians, and cyclists, which pose the risk of collisions. Traditional safety assessment approaches rely on historical crash data, often limited, incomplete, and delayed. Furthermore, they fail to consider the dynamic and stochastic nature of traffic conditions, impacting real-time crash risk.
To address these challenges, proactive safety assessment methods based on traffic conflicts have been proposed. These near-crash incidents offer more frequent and accessible insights into the mechanisms behind crashes. Extreme value theory (EVT) serves as a mathematical tool connecting traffic conflicts to crashes by extrapolating extreme conflict values to the crash domain.
However, existing studies employing EVT for real-time safety evaluation only assess crash risk for the current period, such as the ongoing signal cycle. This approach is still reactive, as it does not provide information about future crash possibilities, which is essential for implementing preventive measures. Therefore, there is a need for a method that can forecast crash probability based on its temporal dependency across signal cycles.
About the Research
In the present paper, the authors introduced a bi-level framework for real-time crash risk forecasting (RTCF) at signalized intersections, leveraging the advantages of artificial intelligence (AI) techniques and traffic conflict data. This framework comprised two levels: (1) a non-stationary generalized extreme value (GEV) model, which in real-time estimated the rear-end crash risk at the signal cycle level utilizing traffic conflict data extracted from video footage, and (2) a recurrent neural network (RNN) model that predicted the crash risk of subsequent signal cycles using past and present input data.
The study applied the proposed framework to three four-legged signalized intersections in Southeast Queensland, Australia. Cameras positioned on 6.5 m high poles near each intersection captured traffic movements during daylight hours over four weekdays (6 am-6 pm). An AI-based automated video analysis platform processed the footage, identifying rear-end conflicts using the modified time-to-collision measure (MTTC).
Additionally, AI techniques like you only look once (YOLO) and deep simple online and real-time tracking (DeepSort) detected and tracked vehicles and pedestrians, extracting their trajectories and conflicts. Loop detector data provided information on signal timing and phasing. Moreover, several covariates, including traffic flow, shockwave area, platoon ratio, and average speed for each signal cycle, were computed using a data fusion algorithm.
The GEV model, developed through Bayesian estimation, facilitated separate GEV distribution estimation for each signal cycle, enabling the calculation of crash risk. Model parameters were parameterized based on covariates to accommodate time-varying factors influencing crash risk. Comparisons between estimated crash frequencies from the GEV model and historical crash records demonstrated a close match.
The RNN model was trained using 70% of the data, validated with 20%, and tested with 10%. It effectively captured the temporal correlation among crash risks across contiguous signal cycles and forecasted future crash risks with high accuracy. Furthermore, the mean absolute error and the relative absolute error of the RNN model were used as the performance measures and showed low values for all study locations.
Research Findings
The authors compared the expected crash frequencies estimated by the GEV model with the historical crash records for the study sites and found a close match between them. They also showed that the GEV model parameters have consistent and intuitive signs, indicating that traffic flow, shockwave area, and average speed have a positive effect on crash risk, while platoon ratio has a negative effect. Additionally, they observed that the GEV model generated separate distributions for each signal cycle, noting significant differences between cycles with positive and zero crash possibility.
Subsequently, the researchers fed the crash risk of each signal cycle estimated by the GEV model to the RNN model and evaluated its performance using mean absolute error and relative absolute error. They found that the RNN model accurately predicted the crash probability for future cycles, with an average relative absolute error of less than 10% for all study sites. Additionally, they tested the model's performance and found that it performed better when trained with more data.
The proposed framework for RTCF has several applications for proactive traffic safety management at signalized intersections. For instance, it can provide real-time feedback to traffic engineers, planners, drivers, and road users, empowering them to adjust signal timing and phasing to mitigate crash risk and avoid hazardous situations. Furthermore, it can facilitate developing and evaluating intelligent transportation systems, like cooperative intersection collision avoidance systems, which communicate with vehicles and infrastructure to prevent accidents.
Conclusion
In summary, the novel method is promising for estimating and predicting crash risk at signalized intersections in real-time, utilizing AI techniques and video data. The authors demonstrated the potential of combining EVT and RNNs to capture the temporal dependency among crash risks across signal cycles, thereby offering valuable insights for proactive safety management.
The researchers acknowledged limitations and challenges and suggested future work could extend the framework to encompass other types of crashes and conflicts, such as angle or vehicle-pedestrian collisions, and to various types of intersections, including roundabouts or unsignalized junctions. Additionally, they recommended validating the framework using more granular and comprehensive crash data and exploring the impact of other factors such as weather, lighting, and road geometry on crash risk.
Improving Urban Mobility with AI Traffic Models