Research On Human Fall Detection Method Based On Improved Joint Point Extraction Algorithm

There are more and more safety hazards in life,especially when they fall,and the number of people injured or killed by falls is gradually increasing every year.To address this problem,an efficient detection method to identify fall actions can greatly reduce injuries caused by falls.In this thesis,we study the techniques related to fall detection based on computer vision class,and the main work is divided into two points as follows.(1)To address the problem that the complex environment in the background of fall detection affects the accuracy rate,this thesis proposes an algorithm for fall detection combining nodal features and fall attribute features,with the following main steps: first locating the human part in the image using the human localization algorithm,then extracting the human skeleton using an improved depth-separable skeleton extraction method,and then using the human skeleton features combined with static and dynamic human fall The human skeleton features are combined with static and dynamic features to determine the human fall behavior.Among them,the static features are the distance between the height of the human body and the ground,and the dynamic features are the falling speed of the center of mass node of the human body,the angle between the main torso of the human body and the ground,and the node positions of the three features of the human body are simulated using the skeleton shape extracted earlier.And the corresponding thresholds are set for these three feature conditions,and the threshold range is used to identify the normal behavior and falling behavior of human body,the proposed method is validated on a relevant dataset and achieves excellent results compared to other methods.(2)To address the problem that confusion of similar behaviors affects the recognition degree,this thesis proposes a pose based two-stream network fall detection algorithm which integrates human pose information and spatial information.A two-stream network model combining deep pose features and spatial features is jointly used,first extracting human pose features using an improved deep separable convolutional network model,obtaining pose features and then using LSTM structure to further learn human pose features to achieve feature refinement,then using CNN network to obtain deep spatial features and refine the spatial features,the refined deep pose features are fused and connected with spatial features at the end of the network to output feature values of the same size scale,and finally the output fusion results are classified and recognized.To verify the effectiveness of the proposed method,experimental tests are conducted on three datasets and compared with other methods,and the results achieved are better than those of other methods.