Multiple States Fall Detection System for Senior Citizens

Since the situation of aging society and injury in elderly have become big problems in our daily life, an experiment which uses several models to detect people's action and selects the appropriate one in fall detection is presented in this paper. This study carries out a series of meticulous classification of user actions. In ideal environment of the experiment, the RealSense D435 depth camera developed by Intel is used to obtain the user's actions videos, and the datasets of images are created according to the action of objects. After binarization, normalization, and recovery, the images are saved as sample database, including training data and testing data for the models. Then the performance of the models are tested after training the models of machine learning (HOG + SVM) and deep learning (AlexNet and GoogLeNet). In these models, a pretreatment layer is constructed to process the images to improve the recognition rate. Finally we analyze the result of each experiment to select the appropriate method, which can perform with the best efficiency and recognition rate.