A Fall Detection System Based on Human Body Silhouette

Elderly care system is one among the most popular research topics in biomedical health-care system design as aging has emerged in different countries. We present a biologically-motivated system to detect unexpected falls in real-time video sequences. The system employs event-based temporal difference image between video sequences as input and extracts static features like aspect ratio and inclination angle of the human body silhouette in unobserved video, which is adopted to improve privacy protection. This method has less computation than those methods using motion dynamic features. Meantime, since time difference is an important factor to distinguish fall incident and lying down event, the critical time difference is obtained from the experiments and verified by statistical results. With the K-Nearest Neighbor (KNN) classifier and the critical time difference, this system presents an accurate approach to detect fall incidents. 86.11% average recognition rate is achieved in the experiment. Compared with other methods of motion dynamic features categorization, our proposed system shows great computational savings, and it is an ideal candidate for hardware implementation with event-based circuits.