Real-Time Traffic Accident Detection From Speed Data: A Visual Method Based on Multi-Perspective Spatiotemporal Tensors

Abstract

Timely detection of traffic accidents is crucial for improving the road safety and traffic efficiency. However, existing methods typically suffer from the reliance of video cameras, coarse granularity, inadequate real-time responsiveness and low accuracy. To address these challenges, this study proposes a novel two-stage spatiotemporal accident detection (TSTAD) approach. By integrating the real-time speed data, historical speed data, similar patterns and speed entropy into a multi-perspective spatiotemporal tensor, the congestion waves caused by accidents are underlined. Subsequently, a two-stage detection framework is designed, which takes the spatiotemporal tensor and auxiliary features (e.g., weather) as inputs. Stage-I classifies that whether accidents happen or not. Stage-II exploits deep visual techniques to detect the occurrence times and locations of accidents. Furthermore, a dynamic sample caching strategy is introduced to speed up the training procedure. Validated on two real-world datasets, TSTAD outperforms baselines by a 3%–10% detection accuracy and over 50% computational efficiency.