Research On Fall Recognition Based On Graph Convolutional Neural Network

With the aging of China’s population becoming more and more serious,falling has become one of the most serious factors threatening the health of the elderly.In most cases,the elderly will lose mobility after falling,which may lead to the loss of the best treatment time due to the failure of timely notification to their families.If the elderly can be found and reported in time after falling,the subsequent harm to the elderly will be greatly reduced.In this paper,a multi-stage human falling behavior recognition framework is proposed by using computer vision method,which successfully realizes the real-time detection of falling behavior.Firstly,the frame detects the position of human body in the video frame by frame,then cuts out the part containing human body and sends it into the human keypoint detection network,which is responsible for detecting the coordinate of human keypoint in the image.Finally,the coordinate of multiple frames of human keypoint is integrated and sent into the fall recognition network for fall detection.The main work of this paper is as follows:A lightweight human position detection algorithm is proposed.The algorithm improved the main feature extraction network of the model,which not only ensured the speed of human position detection,but also improved the detection accuracy.The algorithm improved the feature pyramid structure,and improved the detection ability of different scale targets by integrating the low resolution and high semantic information extracted by neural network with high resolution and low semantic information.A dynamic sample matching method is proposed to avoid the problem that the detection accuracy decreases due to the imbalance of positive and negative samples.A lightweight human keypoint detection algorithm from top down is proposed.The algorithm improves the feature extraction ability of backbone feature extraction network.A training method of random space transformation is proposed,which enables the model to accurately detect the position of human keypoints in the picture when the human body position detection frame is not accurate.A keypoint non-maximum suppression algorithm is proposed to remove redundant human keypoints and improve the accuracy of human keypoint location recognition.An adaptive spatio-temporal graph convolutional neural network is proposed to detect falling behavior.The algorithm can automatically learn the correlation between human bone points by input data.The time convolution structure and graph convolution structure were designed to extract the time and space features of the data respectively to realize the detection of fall behavior.