Study On Video Fall Detection Based On Skeleton And Auto-encoder Models

At present,China has entered the stage of population aging,and the elderly are the most frequently suffered from falls.If the fall event can be detected in time,it can buy time for the rescue.In recent years,with the rapid development of computer vision technology,feasible ways and methods are provided for video-based fall detection.Since fall is an accidental abnormal action,this paper applies the video anomaly detection framework based on the auto-encoder models to fall detection.The main work of this paper is as follows:Firstly,this paper briefly introduces the research status and development trend of fall event detection,and gives the shortcomings of the existing methods,thus clarifying the research content of this paper.Secondly,this paper models the skeleton sequence of the moving target in the video as a skeleton spatial-temporal graph,and describes the spatial graph convolution and the temporal graph convolution.Combining with the design idea of the existing auto-encoder model,a spatial-temporal graph convolutional auto-encoder is proposed for fall detection.Thirdly,aiming at the problem that the human skeleton cannot be completely extracted due to occlusion or large deformation,a skeleton complementation method based on spacetime constraint is proposed.At last,under the condition that the skeleton is completed,the auto-encoder is used for fall detection.For the case that the skeleton cannot be completed,this paper extracts head motion features from the sequence of head joint,and combine it with a composite LSTM auto-encoder,to achieve fall detection.The methods proposed in this paper achieve superior detection results on the public dataset over existing unsupervised methods and the results are close to supervised methods`.It shows that the proposed method provides an effective technical approach for video-based fall detection.