Research On Computer Vision-based Fall Detection

The world’s aging trend is growing, and the people’s habits of modern life have made the most of the elderly live alone, statistics show that falls are the main cause of injury of the elderly. If the falls of older people can not be timely found and implementation of treatment, it will bring more serious consequences. The use of advanced computer technology, sensor technology and image processing technology can achieve fall detection automatically, which can not only provide timely treatment for people that have been injured by falling, but also it can reduce treatment costs, and effectively reduce the possibility of death due to delays in detection and treatment,and reduce the costs of elderly care, improve the quality of life of older persons in the labor.In this thesis, a vision-based human fall detection algorithm is studied, the main work is as follows:First, the analysis of the research background and significance of the human fall detection problem have been taken, then take an overview of the status of the human fall detection at home and abroad as well as the main problems on the main content and chapters framework of this article.Secondly, several vision-based human fall algorithm based on the spatial, temporal characteristics of the human body shape by modeling human fall detection, human fall detection based on inactivity/change of human shape and a gesture-based human fall detection that based on the three-dimensional position of the head is reviewed and analyzed. Eventually their advantages and disadvantages have been compared.Thirdly, the fall detection algorithm based on the depth image of the human body have been analyzed. Kinect is taken as a video capture device to build experimental dataset, we made a large number of fall-detection experiments, and we establish a fall detection dataset based on color video and the depth video. Body contour can be extracted from the depth image by foreground subtraction, then the body’s contour curvature scale space characters is the fall detection features. We take ELM as the classifier to classifier the human’s different activities including falls etc. The sequence of the classifying can give us the detection of falls. Because of the accuracy of extracted body contour from the depth image, and the feature of scale invariance of curvature scale space features, as well as that the speed of ELM training time is short, the proposed fall detection algorithm based on the depth image of the human body has good detection accuracy and it is efficient.Fourthly, most vision-based human fall detection algorithm need to extract contours, which can not be achieved efficiently in the case of parts blocked of the human body, so we propose a new human fall detection algorithm based on deformable model of mixture of multiple components. The algorithm does not require the foreground extraction, it can take fall detection of the human body through the analysis of the posture of the human body. Thanks to the multiple components of the model, it makes the various parts of the object can be used as the characteristics of the effective detection. When the human body part is blocked, it can still be able to effectively detect the body posture, which makes the fall detection algorithm stronger robust and adaptability.Finally, the conclusions are given with recommendation for future work.