Detection of Occlusion-Resistant Based on Improved YOLOv7

Aiming at the problems that crowded objects have serious occlusion and few effective features in complex scenes, an improved YOLOv7 algorithm is proposed to solve the occlusion problem. Firstly, a Focus module is introduced into the backbone network, and the ELAN (Efficient Layer Aggregation Networks) module is modified to the DELAN (deep ELAN) module to enhance the network’s ability to extract effective features. Secondly, by adding P2 layer features in the Neck and modifying the SPPCSPC (Spatial Pyramid Pooling, Cross Stage Partial Channel) module to HSPPCSPC (huge SPPCSPC) module, the feature fusion ability and network robustness are enhanced. Finally, to improve network convergence and detection accuracy, the BCE Loss (Binary Cross-Entropy Loss) has been replaced by the Focal Loss, and the Complete-IoU Loss has been substituted with the Efficient-IoU Loss. Experiments are performed on the Crowd Human dataset. Compared to the original YOLOv7, the improved algorithm shows a 3.22% increase in Recall, a 4.05% increase in mAP0.5test, and a 5.44% increase in mAP0.5~0.95test. The improved YOLOv7 achieves higher detection accuracy while maintaining real-time detection capabilities.