Research And Application Of The Elderly Fall Detection Method Based On Video Surveillance

With the increasing aging of the population,the pension issue will become the focus of social and national attention.In the aspect of personal safety,falling has become an important factor of death and disability of the elderly.In the case of falling caused by illness or accident,timely detection and rescue measures have become an urgent problem to be solved.With the rapid development of artificial intelligence technology in recent years and the continuous improvement of computer performance,more and more researchers pay attention to the detection of human body state through visual means.At present,home cameras and monitoring in various public places have been widely popularized.In the video surveillance scenario,the fall behavior can be detected automatically and the relevant nursing staff can be informed in time,which can solve the problem of elderly safety care with low hardware cost and human cost.Based on the correlative algorithms of object detection and attitude estimation,this thesis studies the human fall detection method,and designs and implements the fall detection system.The main work of this thesis is as follows:(1)This thesis studies the positioning method of human body boundary frame based on YOLOX algorithm,and detects whether the fall.Firstly,relevant fall detection data sets were collected,and the data sets were re-labeled and data enhanced.Secondly,in order to improve the detection effect of YOLOX model,CBAM convolutional attention module and ECA channel attention module are introduced,and the fusion methods of the two modules are proposed to verify the effect of improving the performance of YOLOX network.The experimental results show that the average detection accuracy of the improved YOLOX network reaches 97.24%,which is 1.78% higher than that of the original model.(2)This thesis studies the bone point feature fusion method based on Open Pose algorithm.Open Pose was used to extract the information of human bone points,and a human body boundary frame correction method was proposed in view of the detection errors caused by Open Pose bottom-up algorithm.In order to obtain human motion information comprehensively,a bone point feature fusion method was proposed.The sliding window was used to record the changes in the data of bone points in successive frames.According to the data of bone points in a single frame image,the characteristics such as the relative position of bone points,the Angle of bone point vectors,and the aspect ratio of the external rectangular frame of the human body were defined.At the same time,the characteristics such as the speed and angular acceleration of bone points were defined considering the information of human movement between adjacent frames.In order to reduce the amount of calculation,the feature dimension of bone point data was reduced by principal component analysis,and the fusion feature of bone point was obtained for subsequent fall action recognition.(3)A human fall detection method based on bone point fusion characteristics was studied.This thesis analyzes the principles of KNN algorithm,Adaboost algorithm and SVM algorithm,and uses them as classifies bone point fusion features.The bone point information extracted was used to conduct the fall discrimination experiment and fall direction discrimination experiment.In the fall discrimination experiment,the accuracy of bone point feature fusion method reached 97%,which has obvious advantages compared with other algorithms.In the orientation discrimination experiment,the human fall direction was classified and the risk level was set according to the risk degree of the fall direction.The effectiveness of the bone point fusion feature was proved through the experiment,and the accuracy of the fall direction detection reached 91.5%.(4)The fall detection system is designed and developed.Hardware selection was carried out according to the needs of the system,man-machine interaction was realized through Py Qt technology,and images from surveillance video were extracted for detection.The system includes user login,detection system control,SMS notification and other functions.Once the situation of falling,timely inform the relevant personnel to take measures.In this thesis,fall detection research is mainly conducted through the way of computer vision,combining the YOLOX algorithm and Open Pose algorithm to extract the fall features,on the basis of fall discrimination,fall direction detection,design and implementation of the fall detection system,providing a new scheme for the development of the elderly intelligent monitoring system.