Sudden fall classification using motion features

Monitoring of abnormal activities under video surveillance research area is important due to providing comfort and safety living for the society. The popular scenario is to learn pattern of normal activity, and subsequently detect abnormal events in the scene. Instead of detecting abnormal event, we propose to model the sudden change in the event specifically fall event that deviates from the normal activities. We learn the motion features namely, motion history histogram (MHH) and motion geometric distribution (MGD) across image in the frame sequence. Then, we propose a classification strategy using biological inspired feedforward network that can detect sudden abnormalities in the event. We test the algorithm on real dataset and found that our approach is able to distinguish the transition state between walk and fall.