Improved Alphapose Combined With GCN For Human Fall Detection Model

Detecting and alerting the elderly in time when they fall can effectively protect their lives.In this paper,we propose a real-time fall detection framework that combines target detection,target tracking,human pose estimation and behavior recognition techniques based on deep learning and computer vision.The framework first performs human target detection for each video frame of the video to extract all the human detection frames in the current frame,then uses target tracking techniques for human tracking to get more stable and smooth human detection frames,next performs single person pose estimation for each tracked human target to extract the human skeletal key points of each target,and finally uses human behavior recognition techniques for all The skeletal keypoint maps are finally processed by human behavior recognition techniques to classify the human state and thus determine whether a fall has occurred.This paper mainly does the following work.(1)To address the problem that the top-down human posture estimation method is limited by the target detection results,the optimized YOLOX＿s target detection algorithm is used instead of the original YOLOv3 target detection method in Alphapose as an improvement to obtain more accurate human detection results.The experimental results on Crown Human dataset show that the optimized YOLOX＿s in this paper has certain advantages in real-time,accuracy,number of parameters and other evaluation indexes.(2)Combining YOLOX with Deep Sort target tracking algorithm to achieve real-time multitarget human tracking and to make the resulting human detection frame more stable and smooth.(3)A human pose estimation algorithm combining Swin-Transformer sliding window multi-head attention and SENet channel attention is proposed for the problems of partial human body occlusion and inaccurate human skeleton key point extraction.The experimental comparison with Open Pose and Mask-RCNN algorithms on COCO2017 pose dataset shows that the optimized human pose estimation algorithm in this paper can extract higher quality human skeletal key points.(4)A human behavior recognition algorithm based on graph convolutional neural network and skeletal key point information is proposed for fall detection.Experimental comparison with other fall detection algorithms on the Le2 i dataset shows that the fall detection algorithm in this paper has a high accuracy rate of 92.2%.The experiments show that the improved Alphapose combined with GCN for fall detection can perform fall detection in real time with high accuracy and have good detection for partial limb occlusion,which has high practicality.