Research On Human Fall Detection Based On Skeleton Information In Home Scene

Falls can cause serious consequences such as fractures,brain damage,and coma,and are an important cause of accidental injuries to humans.In addition,falling can also cause people to lose confidence in normal walking and create a fear of falling,which can lead to a greatly increased incidence of depression.Human behavior in the home environment is complex and changeable,and there are a large number of behaviors that are easily confused with falls,which put forward higher requirements for detection methods.Therefore,accurate and timely detection of falls is of great practical significance to safeguard the physical and mental health of the human body.The main work of the paper is as follows:(1)Design and collect a multi perspective human behavior fall dataset,and analyze the impact of feature vectors composed of skeleton joint point information on fall behavior detection results from different perspectives..Aiming at the problem of non-uniform feature vector thresholds caused by multi-view data,based on machine learning methods,this paper proposes a multi-threshold classification fall detection algorithm to reduce the impact of errors from different perspectives on fall detection.The accuracy of the algorithm is It is 92.59%.(2)In view of the high misjudgment rate of machine learning methods and the inability of ST-GCN to fully mine the dynamic information and context dependencies neglected in feature mapping,the advantages of remote and relational modeling based on the self-attention mechanism,This paper proposes a fall recognition method based on graph convolution and self-attention mechanism(Graph Convolution Self Attention,GCSA)to better mine context dependencies and temporal features,thereby improving human fall behavior recognition in complex home scenes the accuracy rate.A large number of experiments on the self-built data set have proved that the GCSA model proposed in this paper has achieved higher recognition accuracy.The recognition accuracy of the GCSA model is 95.26% under the mixed perspective.At the same time,the network uses the skeleton information as the input data with a small amount.And it is robust to environment and illumination changes.(3)In view of the inaccuracy of extracting the skeleton information itself and the misjudgment of confusing actions,this paper proposes a two-stream detection and recognition network based on skeleton information and optical flow information(Two-stream Algorithm Based on GCSA,TS-SAGC).The network combines skeleton information and optical flow information weightedly to realize human fall behavior detection in complex home scenes.Adding the optical flow branch can not only effectively make up for the shortcoming of only using the skeleton information as a single-stream input for action recognition,but also more accurately capture the instantaneous acceleration information when the human body falls.The fusion of the two improves the accuracy of the fall detection algorithm,and the fusion of the skeleton The accuracy rate of the dual-stream detection and recognition network of information and optical flow information is 98.23%.