Improved Pedestrian Fall Detection Model Based on YOLOv5

Falls are an important factor in the death and injury of workers in complex operating environments. In view of the problems of missed detection and wrong detection of the original YOLOv5 network, this paper proposes a pedestrian fall detection model based on YOLOv5. The self-built pedestrian data set is used for fall detection research. In order to weaken the interference of complex background on network feature extraction, an improved SENet attention mechanism is proposed, which helps the network to pay more attention to the fall posture. In addition, in order to reduce the missed detection rate, Soft-NMS is introduced to replace the original NMS of YOLOv5. The results show that the mAP of the fall pedestrian detection training set of the improved model is increased from 97.62% to 98.33% which proves that the improved model can better meet the requirements of pedestrian fall detection than the unimproved YOLOv5.