Research On Fall Detection In Indoor Surveillance Video Under Visual Privacy Protection

With the aggravation of population aging and the increasing common of empty nest phenomenon,how to effectively supervise and protect the health of the elderly living alone has become a hot issue of the whole society.Surveys have shown that accidental falls are the main cause of injuries and even deaths in the elderly,and this problem is particularly obvious among elderly people living alone.For a long time,the lack of reliable fall detection system has been a practical problem for home monitoring.Therefore,designing a reliable and effective fall detection system has far-reaching significance and social value.Among the existing fall detection methods,fall detection algorithms based on computer vision are widely used due to their low cost,easy maintenance of equipment and low interference.The realization of fall detection based on computer vision relies on video surveillance.Although video surveillance can effectively monitor the behavior of elderly people living alone at home,there is a contradiction between this traditional video surveillance method and personal privacy.Generally,people do not want to be monitored at home all the time.The high-definition camera contains a wealth of information,which also includes life privacy in addition to behaviors that require attention such as falls.In order to balance fall detection and privacy protection,this paper proposes a new fall detection method with visual shielding function to ensure the home safety of the elderly while protecting their personal privacy.Starting from the visual shielding method,this paper designs a coding model based on multi-layer compressed sensing,so that the encoded video has a visual shielding effect.Based on the analysis of the data characteristics of compressed video,the key technologies of video feature extraction and classifier design are studied.The research work of this paper mainly includes the following parts:(1)This paper analyzes the research status and common methods of traditional fall detection algorithm,and focuses on several common fall detection methods,as well as the feature extraction algorithm and common classifier related to fall detection.(2)For the visual privacy protection of video,based on the compressed sensing(CS)theory and the idea of image block compressed sensing,the single-layer block compressed sensing process is extended to multi-layer,and a multilayer compressed sensing for visual shield sensing coding(MCSVSSC)model is proposed.In addition,in order to make the compressed video not only have the effect of visual privacy protection,but also retain the video information with high fidelity,this paper compares the commonly used measurement matrix and selects a new generation mechanism of measurement matrix.The MCS-VSSC model not only realizes visual shielding,but also greatly reduces the amount of data processed by the video,laying the foundation for subsequent intelligent recognition.(3)By analyzing the characteristics of the MCS-VSSC video data,the spatiotemporal feature suitable for compressed video data are proposed.On the basis of multilayer CS processing,the visual information of MCS-VSSC video is weakened.Traditional spatiotemporal feature extractors have certain requirements for image resolution,so they are no longer suitable for the framework of this paper.However,no matter whether the image resolution is reduced or not,the local binary pattern on three orthogonal planes(LBP-TOP)can get the feature description by finding a corresponding relationship between pixels in the video.In addition,because the measurement matrix used in this paper can effectively retain image information,and the timing information between videos can also be effectively retained after the same observation processing,this paper uses LBP-TOP as the behavior feature expression in video.What is more,in order to enhance the features of the moving regions in the video and weaken the surrounding information and noise,this paper proposes a weighted method based on the motion contribution,which is called block weighted local binary pattern on three orthogonal planes(BWLBP-TOP).(4)In order to further improve the recognition accuracy of compressed video,a private information-embedded model based on a generative adversarial network(GAN)is proposed.Although the proposed BWLBP-TOP feature can describe the behavior of MCS-VSSC video well,the characteristics of MCS-VSSC video data lead to the difference between the spatiotemporal features of MCS-VSSC video and that of the original video,which will directly affect the subsequent behavior recognition.In order to enrich the spatiotemporal features of compressed video and improve its recognition performance,this paper is inspired by the idea of GAN,and proposes a model to make compressed video learn the original video features to enrich its own feature expression,which is called private information-embedded model.Experiments on two common fall datasets show that the proposed spatiotemporal feature extractor and classifier have good performance.