Pedestrian Fall Detection Using Improved YOLOv5

Abstract

An end-to-end fall detection method was developed using the YOLOv5s model to accurately locate a person and monitor their fall behavior in a crowd. We added the SE attention mechanism to the second and fourth CSP1_X structures in the network using feature extraction to locate a target more precisely. The spatial pyramid pooling and fully connected spatial convolution (SPPFCSPC) structure was designed to replace SPP to extract the information of the target in different scales effectively and enhance its feature expression ability and detection accuracy. Compared to the previous model, the precision, mean average precision (mAP), and recall rate of the YOLOv5s-2nd-4th-C3SE-SPPFCSPC model increased by 3., 6.2, and 2.9%, respectively. the mAP of the fall category increased by 7.3%. The developed model showed improved detection ability which surpassed that of the original YOLOv5s model.