Detection of Human Fall in Video Using Shadow Information

This paper presents a novel algorithm for detecting human fall incidents in video clips. This algorithm can effectively differentiate between fall-down and fall-like incidents such as sitting and squatting. Normally, complex 3-D models are required to solve this issue. However, to reduce the high computational cost, we use the shadow information instead. Furthermore, this algorithm is able to efficiently work under bird's-eye view camera setting. The experimental results show that the proposed shadow-assistant approach can achieve a high detect rate and low false alarm rate from very short frame sequences, 1-10 frames, while satisfying real-time constraints.