Fall Detection Based On Human Pose Recognition

As the population aging problem becomes serious,accidental falls have become a serious threat to the healthy life of the elderly,and the study of fall detection has high social significance.Based on the computer vision method,this paper proposes a multi-stage fall detection framework that combines human target detection,human posture estimation and action recognition.First,detect all the human body bounding boxes in the video or picture,and then use single-person pose estimation to recognize each person's body skeleton image,and finally use motion recognition technology to classify all body skeleton images to determine whether they fall.The main work of the thesis is as follows:(1)Propose an occluded human detection algorithm based on hybrid attention mechanism.In order to solve the common occlusion problem in human detection,the channel attention mechanism is added in the feature extraction process to change the attention weight of different feature channels of the network,and the spatial attention mechanism is added to change the attention weight of the pixels in the feature map,through the attention weight.In the process of human target detection,focus on the part of the human body that is not blocked,so as to achieve the purpose of detecting the blocked human body.The result of verification on the test set shows that the algorithm can solve the human body detection problem that is blocked by obstacles in the human body detection process.(2)A top-down human pose estimation algorithm is proposed.Based on the detection of the bounding box of the human body based on the occluded human body detection algorithm based on the hybrid attention mechanism,the detection of the bounding box of the human body may have inaccurate positioning and lead to deviations in the position of the human body.The spatial transformation network is used to determine the bounding box of the human body.Coordinates undergo affine transformation to correct their position.The results of verification on the test set show that the algorithm can effectively and correctly estimate the visible parts of the human body posture,and reduce the influence of obstacle occlusion on the human posture estimation.(3)A human fall detection algorithm based on spatio-temporal graph convolutional network is proposed.Based on the detection of the human skeleton map based on the human pose estimation algorithm,the time information is lost when the ordinary 2D convolution extracts features from the video,and the spatio-temporal map convolution is used to extract the feature information from the two parts of the time dimension and the space dimension.The result of verification on the test set shows that the algorithm can effectively detect a fall when the human body is occluded,reducing the impact of occlusion on human fall detection.Experiments have proved that the occluded human detection algorithm based on the hybrid attention mechanism can better solve the obstacle occlusion problem in human detection.The m AP reaches 52.18%.The human pose estimation algorithm can correctly estimate the visible part of the human pose,Kps AP Reaching 69.7%,the multi-stage human fall detection framework can detect whether a human body blocked by an obstacle falls in real time,which has good practicability.