Research On Human Contour Acquisition And Feature Extraction Algorithm In Fall Detection

Due to the aging of the population,the incidence of falls in the elderly is increasing year by year,many scholars began to involve the field of fall detection.Traditional detection methods usually use a single sensor,or use image processing technology to obtain the target information,to judge the fall behavior.With the deepening of technology learning,many kinds of sensor combination,deep learning and other technologies have gradually become the mainstream.This dissertation include two kinds of sensors,accelerometer and camera,they are used to collect the target behavior data,analyze the acceleration and image information,select the feature method that can distinguish the fall and non-fall to build the feature vector,and use SVM(Support Vector Machine)to detect the fall behavior to achieve multi-sensor technology.This paper focuses on the processing of video data,that is,foreground target extraction and feature extraction,mainly including the following contents:(1)Building human behavior database.The system uses two ways to obtain the data features: one is to wear the detection device,and use the accelerometer to obtain the acceleration features of different behavior states in real time;the other is to collect the image data through the fixed camera,and build the data base of human body in various states.Combining the characteristic data of two kinds of sensors to analyze and predict the state of falling and non-falling.(2)Contour extraction.Firstly,preprocess the collected video data,then compare and observe the contour extraction renderings to confirm the extraction algorithm.In addition,the shadow area is located by transforming to HSV space,and then the shadow part is removed by "and" operation with the foreground,so as to solve the problem that the extracted feature data is not suitable for fall detection.(3)Feature extraction.In addition to the relative height of the center of mass,the long axis of the ellipse,the aspect ratio and other methods,this paper introduces the geometric section line and the central line features,analyzes the characteristicwaveform of the video target's walking,jumping,lying down and falling behaviors,and compares the applicability of the feature method.(4)Fall detection.The acceleration feature,image feature and fusion feature vector group are constructed,and SVM is used to determine the fall.In addition,the effects of geometric intercept and central line feature on the test results are compared.The system designed in this paper includes data acquisition,receiving,contour extraction,feature extraction and fall classification module,which can ensure the real-time and stability of the system.Compared with the traditional judgment method based on threshold value,this paper adds the method of information acquisition and feature extraction.Through the reorganization of different features,the feature group with the best accuracy is constructed,the false alarm rate of the system is reduced,and the accuracy rate of fall and non-fall behaviors is improved.