Intelligent Video Surveillance for Monitoring Elderly in Home Environments

In this paper we propose a novel method to detect and record various posture-based events of interest in a typical elderly monitoring application in a home surveillance scenario. These events include standing, sitting, bending/squatting, side lying and lying toward the camera. The projection histograms of segmented human body silhouette are used as the main feature vector for posture classification. k-nearest neighbor (k-NN) algorithm and evidence accumulation technique is proposed to infer human postures. With this technique we have achieved a robust recognition rate of above 90% and a stable classifier´s output. The modified classifier structure also improves greatly the recognition rate of lying toward the camera events as compared to the result of classifier´s structure in GHOST (Haritaoglu et al., 1998). Furthermore, we use the speed of fall to differentiate real fall incident and an event where the person is simply lying without falling.