Research Of Fall Detection And System Design Based On Depth Video And Wearable Sensors

With the evident tendency of aging population, the proportion of old people gradually increases. The phenomenon that elderly people live alone has become a widespread social problem, and the elder’s health and medical issues become social concerns. Physical damage has threatened the elder’s health, and the fall events cause most of it. So, if the fall events can be detected correctly by the informationalized technology, the old people will receive treatment in time, and the caused damage will be minimized.Among the traditional fall detection algorithms, those based on ordinary 2D video are impacted heavily by environment, they are suffering from background interference, foreground extraction difficulties. The others based on three-axis acceleration sensor are suffering from poor stability, and their sensitivity and specificity cannot be guaranteed at the same time. Along with the development of depth sensor, the algorithms based on depth sensor overcome part of the problems mentioned above, but their application scenarios are limited. To fix those problems, a novel fall detection algorithm based on the human skeleton model and wearable device sensors information is proposed. The proposed algorithm is robust, and overcome many problems such as partially blocked issue. The main contribution of this paper as below:(1) A completed fall data base is built in this paper. We consider carefully about the variable of the fall event, and define a fall event set and a daily activity set. There are 660 samples in our data base: 380 positive samples and 280 negative samples. Every sample is made of depth video data and wearable device data.(2) Combined with floor detection algorithm, the fall detection algorithm in this paper uses human skeleton model information to extract reliable feature. Using SVM as classification algorithm, ultimately, 98.211% in sensitivity and 97.464% in specificity is achieved. At the same time, different from the traditional method that uses acceleration sensor data, this paper presents a three-dimensional movement model of wearable device(such as phone).An acceleration sensor, a gravity sensor, and an angular velocity sensor are used to simulate phone’s state for extracting effective features. Eventually, based wearable sensor data, we obtain 94.053% in sensitivity and 94.858% in specificity.(3) In this paper, the depth video information and the wearable device information are fused on two levels: 1) on the feature fusion level, the PCA algorithm is used to fuse the features in two channels. Ultimately, 98.658% in sensitivity and 97.643% in specificity is achieved, which is better than the algorithm which use single channel. 2) On the decision fusion level, the algorithm evaluates information from each single channel. Then we use different fall detection algorithm for different scenario, so that our algorithm can adapt to a variety of fall scenes.(4) In this paper, combined with the proposed fall detection algorithm, a fall detection system is designed and implemented. The fall detection system designed in this paper is scalable, user-friendly, high stable and real-time. Experiments show that the system can adapt to a variety of fall scenes.