考虑跨空间特征重构的行人过街动作检测方法OA

Pedestrians,as vulnerable road users,face elevated risks in traffic environments where hazardous crossing behav-iors constitute a primary contributor to road accidents.Aiming at the problem of action feature extraction,inefficient feature fusion and feature information loss caused by multi-scale and occlusion of pedestrians under the monitoring perspective,a pedestrian's crossing behavior detection network(PCBDNet)based on cross-spatial feature reconstruction is proposed.The backbone network uses a cross-space parallel subnetwork feature structure to enhance the network's ability to learn multi-scale pedestrian features.The neck network fuses features using spatial and channel reconstruction convolutional modules to reduce redundant computation and facilitate the learning of pedestrian action features,and constructs feature enhancers to separate features and perform cross-spatial reconstruction to reduce information loss due to occlusion.The head network uses repulsion loss to improve the coordinate loss function to further improve occlusion target detection accuracy.Experiments show that the mAP@0.5 reaches 89.6%,the mAP@[0.5:0.95]reaches 72.4%,the model size is moderate,and the FPS remains high.The feature visualization experiments show that PCBDNet pays more attention to pedestrian crossing actions,which provides a new solution to enhance the risk reduction of traffic accidents.