A WEARABLE SENSOR BASED APPROACH TO REAL-TIME FALL DETECTION AND FINE-GRAINED ACTIVITY RECOGNITION

摘要

We present a real-time fall detection and activity recognition system that is inexpensive and can be easily deployed using two Wii Remotes worn on human body. Continuously 3-dimentional data streams are segmented into sliding windows and then pre-processed for removing signal noises and filling missing samples. Features including Mean, Standard deviation, Energy, Entropy, Correlation between acceleration axes extracted from sliding windows are trained the activity models. The trained models are then used for detecting falls and recognizing 13 fine-grained activities including unknown activities in real-time. An experiment on 12 subjects was conducted to rigorously evaluate the system performance. With the recognition rates as high as 95% precision and recall for user dependent isolation training, 91% precision and recall for 10-fold cross validation and as high as 82% precision and recall for leave one subject out evaluations, the results demonstrated that the development of real-time, easy-to-deploy fall detection and activity recognition systems using low-cost sensors is feasible.