Research On Fall Detection Method Of Elderly Living Alone Based On Computer Vision

The fall event has become one of the important factors that threatening the health of the elderly.the fall usually cause physical injury or even death to the elderly.With the continuous acceleration of the aging process in our country,it won't be guaranteed the safety of the elderly living alone.If the elderly falls at home alone,it will lead to accidents occurring because of untimely rescue and other reasons.Meanwhile,the development of computer vision makes it have significant advantages and characteristics in many fields.Therefore,it is significant and valuable that researching on fall detection methods based on computer vision.In the thesis,it is researched that fall detection methods based on computer vision,including detection of moving target,fall detection method based on manually extracted features,fall detection method based on automatically extracted features,and test verification based on experimental platform.The main research contents are as follows:(1)Analysis of moving target detection algorithmIt is compared and analyzed of the common moving target detection methods and the principle of the ViBe algorithm,and it is verified that the detection effect of these methods on moving target by experiments.Due to the shortcomings and limitations of common target detection methods in detecting target integrity and adaptability to background changes,the ViBe algorithm has unique advantages in target detection.Therefore it is determined that the ViBe algorithm is used to detect moving target.(2)Research on fall detection method based on multi-feature analysisIn order to reduce the error rate of fall detection,combining with ViBe algorithm and common fall features,a fall detection method based on multi-feature analysis is proposed.First the distance from the center of mass of the human body to the horizontal plane is used to determine whether the moving target is in a standing or normal walking state.If the moving target isn't in a standing or normal walking state,then the aspect ratio of the human body is used to distinguish common daily activity behaviors and fall behaviors.Because of the possibility of misjudgment at this time,the speed of the change in the center of mass of the human body is finally used to distinguish between daily activities behaviors similar to fall and fall behaviors.And it is proved that the effectiveness of the method by experiments.(3)Research on fall detection method based on PCANet and SVMAiming at the deficiency of manual extraction of features on the background environment and the limitation of the requirements for setting the scene in practical applications,it is proposed that a fall detection method based on PCANet and SVM.First PCANet is used to extract image features;and then two SVM classifiers are trained by the extracted features,which are used to obtain frame image labels and video label sequences respectively;finally,the test verification is carried out,after the foreground detection and prejudgment processing,the test video samples are input into two SVM classifiers in turn,and the correctness of the classification is determined by comparing whether the predicted video label sequence is consistent with the set label sequence.Experiments show that this method can accurately distinguish between daily activity behaviors and fall behaviors.(4)Construction of experimental platformThe experimental platform is built by using Raspberry Pi,SIM800C,USB cameras,and test experiments are conducted by real-time monitoring sample videos.It can not only accurately detect the occurrence of fall event,but also automatically send a text messages or make a call when a fall event occurring.It is verified that the feasibility of the proposed fall detection methods in practical application.On the basis of the existing fall detection methods,fall detection methods based on computer vision is made improvements and attempts in the thesis.And based on this,the experimental platform is built to test and verify,which can not only effectively distinguish behaviors fall and non-fall behaviors,but also can inform the guardian at the first time when a fall event occurring.It is hoped that providing some reference for the implementation of fall detection in real life scenarios.