Bridging the Adaptability Gap in Machine Learning Driven Road Safety Systems: A Research Roadmap Toward Safer Streets

ML-based road safety systems also have the potential of reducing traffic accidents and improving pedestrian safety. However, widespread adoption is a major barrier: rigidity. Urban environments are complex and volatile, with changing traffic volume patterns, evolving infrastructure, diverse weather and unpredictable human activities. Modern machine learning models sometimes have difficulty generalising across spatiotemporal variations, resulting in reduced performance when applied outside of controlled or static testing scenarios. In this work we first introduce the adaptation gap that quantifies the discrepancy of a model’s performance in training and its robustness to operational variability in the real world. We are aiming for an ambitious agenda of research filling the gap by advancing generalized models, supporting continual learning and managing context, as well as incorporating human-in-the-loop approaches. The paper puts forward a multilayer perspective for future responsive, resilient and just road safety. It accomplishes that by combining the latest developments in machine learning, computer vision, edge computing and transportation systems engineering. We conclude with ethical considerations and policy implications, advocating interdisciplinary collaboration for improving the safety of urban mobility.