Application Research Of Privacy-Preserving Fall Detection Via Chaotic Compressed Sensing And GAN-Based Feature Enhancement

With the improvement of living standards and the continuous improvement of medical and health conditions,the life expectancy of the population continues to extend,making the problem of population aging increasingly prominent.How to ensure the health of the elderly in their later years has become a hot issue in today’s society.The survey found that the biggest contributor to unintentional injury or even death among older adults is falls.At the same time,because children cannot stay with their parents all the time,when the elderly fall,once they are not found in time,there will be extremely serious consequences.Therefore,providing real-time safety monitoring has important research significance for the quality of life and even life security of the elderly.With the widespread use of video surveillance,video-based fall detection has a good development prospect due to its rich monitoring information,non-contact monitoring methods,and low cost.Therefore,timely and accurate automatic detection of fall behavior in home video surveillance has become a research hotspot in the field of computer vision.Reliable vision-based fall detection systems can not only effectively ensure the safety of seniors living alone,but will also play a very important role in future healthcare and assistance systems.Traditional vision-based fall detection methods are aimed at clear videos,that is,using image processing technology to extract and analyze the characteristics of subjects in high-quality video sequences to determine whether there is a fall phenomenon in the video,but it emphasizes personal privacy.Today,traditional vision-based fall detection methods have serious shortcomings such as personal privacy exposure.In this thesis,the fall detection system based on computer vision,from the two aspects of fall behavior detection and privacy protection,can not only protect the safety of the elderly,but also avoid the privacy leakage of the elderly.Firstly,the original video data is processed at low resolution by using multi-layer compressed sensing technology,so that the processed video can achieve the goal of privacy protection;for the purpose of fall detection.The main work of this thesis can be summarized as the following parts:(1)The main research status and common methods of current fall detection are analyzed,and the commonly used fall detection methods,foreground moving target extraction methods and commonly used behavior classifiers are introduced in detail from three aspects.(2)In order to realize the privacy protection of video data,based on the traditional compressed sensing theory,this thesis introduces the construction of the measurement matrix in detail.On this basis,it analyzes and compares the characteristics of the currently commonly used measurement matrix,and innovatively proposes a chaotic pseudo-random-based Bernoulli measurement matrix.The original video is processed by compressed sensing(CS)using this measurement matrix,and the video data in the visual privacy protection state(VPP data)is obtained.The acquired VPP data can greatly reduce the amount of video processing data on the basis of realizing the function of visual privacy protection,which is convenient for subsequent processing.(3)After the video data is processed by multiple layers of visual privacy protection,not only will some information be lost,but also a lot of noise will be introduced,making it difficult to distinguish the detected target and increasing the difficulty of subsequent fall behavior recognition.Therefore,it is necessary to remove background information.and noise.This thesis first introduces the relevant theory of low-rank sparse decomposition,and through its existing problems,leads to the Generalized Nonconvex Robust Principal Component Analysis(GNRPCA)algorithm used in this thesis,which is used in the specific application of this thesis.More abundant foreground moving target information can be obtained,which provides better feature support for subsequent research.In order to more accurately identify the falling behavior in video data on the basis of visual privacy protection,this thesis starts from the GAN network,relies on the basic architecture of AC-GAN(Auxiliary classifier GAN),and improves the network on this basis.As a new information transfer tool,the improved-AC-GAN architecture enables competitive learning of features between VPP data and original video data,thereby enriching the expression of VPP data’s own features.The experimental results on three open fall datasets show that the method is not only effective in detecting falls in videos,but also has high accuracy.