Preemptive crash risk reduction through a real-time cost-based safety prediction model (RECOSAM) for traffic signal control

This paper proposes a novel real-time cost-based safety prediction model (RECOSAM) and incorporating it in intersection traffic signal control optimisation, complementing the recent advances in deep reinforcement learning (RL)-based adaptive traffic signal control (ATSC). The primary contribution is the development of RECOSAM, a model designed to predict traffic safety risks one step ahead of time, for various signal phase configurations at intersections. The proposed model offers a dynamic safety evaluation strategy, estimating near-future safety metrics for seamless integration into machine learning-based ATSC systems. Extensive experiments validate the model’s effectiveness, demonstrating its potential for adaptive adjustments to mitigate impending safety risks. Perhaps more importantly from an operational policy perspective, the proposed model is capable of finding an optimal and justifiable trade-off between the efficiency of traffic flow and its safety in real-time. A case study showcases the integration of RECOSAM into deep RL for green time optimisation. Results suggest that extended dedicated right turn phases may reduce safety risks, while overly protected phases could lead to inefficiencies in green time allocation and increased congestion. The model’s adaptability across different scenarios is further illustrated, showing its capability to evaluate critical trade-offs between safety and efficiency especially for vehicles trying to make a right turn by finding gaps through traffic coming form the opposing direction (in left-hand-side driving countries—same applies for left turns in right-hand-side driving countries).