Abstract
Accident detection in surveillance video can shorten emergency response times, but a practical detector must also raise few false alarms in normal traffic. This study evaluates an interpretable pipeline that detects vehicles with YOLOv3, tracks them with MOSSE, and raises an alarm when a pixel-based motion-change score exceeds a fixed threshold. On 15 videos that all contain accidents, the pipeline detected 60% of the accidents. However, on the public TAD benchmark, which also contains normal traffic, its score did not separate accident videos from normal videos (test AUC 0.48), and it raised alarms in 14 of 16 normal videos. Controlled experiments traced this failure mainly to tracking errors and the pixel-based score. We therefore propose a configuration that detects vehicles with YOLOv3, tracks them with ByteTrack, and scores each vehicle by its speed just before it stops suddenly. It raised the test AUC to 0.88 and the F1 score from 0.67 to 0.83. Its ranking performance was statistically indistinguishable from that of a fine-tuned video transformer (AUC 0.84), although it needs no training on accident videos and identifies the vehicle and frame behind each alarm. Because false alarms remain frequent, the method is suited only to pre-screening by human operators.