Research On Human Fallins Detection Combining Deep Learning Features

With the development of society,the world's aging population is constantly increasing,so that the health of the elderly become one of the concerns of the community.In the issue of affecting the health of the elderly,fall injury is the most damaging problem.This paper is based on deep learning to study the daily life's human fall detection.First,by comparing the human body detection methods in video sequences and in the picture,the foreground extraction method is selected to carry on the initial detection of the moving human body.Then,the high-level feature extracted by the deep learning framework is combined with the image-based low-level feature to represent the state of human motion.For the judgment of human fall,we construct a cascade support vector machine model.The first layer of support vector machine determines the human movement's state at a moment,and the second layer of support vector machine uses the movement's states in a continuous period of time to determine whether a fall event has occurred.Finally,according to the problems that may arise in real life,the detection method of combining the human descriptor with the classifier is introduced,and the detection ability of the moving body in the complex state is improved by using the PCANet and the Gradient Boosting Decision Tree.In this paper,the human fall detection model obtained 97%sensitivity and 95%specificity in the dataset collected by ourselves,and 93.7%sensitivity and 92.0%specificity were obtained in the public fall data set.