Diffusion Edge Guidance Network and Class Specific NMS for Vulnerable Target Detection in Surveillance Video of Urban Roadway

Accurate detection of vulnerable targets in urban road surveillance videos is crucial in intelligent transportation systems widely used in urban roads. Since the features of vulnerable targets conflict, it is difficult for the target detection algorithm to use the complete target features fully. Therefore, it is imperative to fully mine the features of the feature conflict targets for the detector design. Although there have been methods to enhance the integrity of target features based on target edge features, the edge’s ability to emphasize the target’s internal regions is limited, and there is no guarantee that the emphasized features must be internal features of the target. To this end, we propose DEGNet, which emphasizes the internal features of the target by diffusion and superposition of the target edge in the eight neighborhoods and designs DEAM to guide the target detection network to pay more attention to the internal features of the target from the feature level. In addition, we propose CS-NMS, which adaptively adjusts the degree of influence of the category on the screening of detection boxes. We constructed an urban road vulnerable target detection dataset SC8D with feature conflict scenes and tested the performance of the proposed method in this dataset and real-world scenarios. The results show that our method outperforms existing methods and can effectively apply to fragile object detection in feature conflict scenarios.