Exploring Spatiotemporal Relational Learning With TimeSformer for Identifying the Severity of the Road Accidents

Road traffic crashes represent a major global public health concern, resulting in millions of fatalities and injuries annually. The burden is particularly severe in low- and middle-income countries, with India experiencing a notable rise in traffic accidents—underscoring the urgent need for improved safety measures and emergency response systems. While recent advances in computer vision and deep learning have enabled real-time accident detection, the prediction of crash severity remains largely unexplored. This article introduces a novel task of crash severity prediction aimed at enhancing emergency response and optimizing resource allocation. We present <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Collision Chronicles</i>, a curated dataset of 1500 annotated dashcam videos specifically designed for severity classification. Our proposed architecture combines a TimeSformer backbone for spatiotemporal feature extraction, a bidirectional long short-term memory (LSTM) for sequential modeling, and multihead attention (MHA) for contextual refinement. We also introduce a custom accuracy metric that rewards contextually appropriate predictions, reflecting the ordinal nature of crash severity. Furthermore, we extend our framework to binary accident detection using a balanced dataset of accident and nonaccident videos. To support edge deployment, we apply posttraining dynamic quantization, reducing model size by 3.6<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\times$</tex-math></inline-formula> without compromising predictive accuracy. These contributions establish a robust foundation for real-time crash analysis, scalable deployment, and proactive road safety interventions.