Highway spillage detection using an improved STPM anomaly detection network from a surveillance perspective

Abstract

Spillages may cause traffic congestion and incidents and seriously affect the efficiency of traffic operation. Due to the changeable shape and scale of a spill on a highway, the location of the spill is random, so the current background extraction and object detection methods cannot achieve good detection results for the spill. This paper proposes a highway spill detection method using an improved STPM anomaly detection network. The method is based on the STPM network and achieves detection through FFDNet image filtering, calculation of the global correlation features of the student and teacher networks, contour positioning of spillages in the feature map, and automatic collection of positive samples to train and update the model, achieving high-precision identification and positioning of the spillages. The experimental results of the custom-built top-view road surface spillage dataset and the MVTec anomaly detection dataset show that the method proposed in this paper can obtain an AOC-ROC value of 0.978 and a PRO score of 0.965 and can distinguish between spillages and reflective cones, avoiding the problem of false detection when spills are similar in appearance. Therefore, the proposed method has value in the research and engineering application of spill detection in special highway scenes.