AI in Accident Prediction: A Review of Crash Risk Modelling, Spatio-Temporal Deep Learning and Dashcam Accident Anticipation

Road traffic crashes kill more than a million people each year, and India records one of the highest numbers of road deaths in the world. Predicting where, when and how accidents are likely to occur allows authorities to deploy enforcement and engineering measures, and allows vehicles to warn drivers before a collision. This paper reviews artificial intelligence for accident prediction at two scales. At the network scale, it examines statistical crash-frequency models, machine learning comparisons with traditional regression, and spatio-temporal deep learning that combines historical accidents, traffic, weather, road features and points of interest to estimate accident risk for each region and time. At the vehicle scale, it reviews accident anticipation from dashcam video, in which recurrent and relational models attend to road agents and estimate the probability of an imminent collision several seconds in advance. Datasets, evaluation metrics and explainability methods for identifying risk factors are discussed. A five-layer architecture comprising data sources, data integration, risk modelling, explanation and intervention layers is proposed, together with research gaps, objectives, a comparison of approaches and challenges such as under-reporting, class imbalance, data sharing and false warnings. The review concludes that AI-based accident prediction, linked to concrete interventions, can contribute significantly to road safety.