[Illinois] Recent Advances in Kinect-Based Action and Event Recognition

The emergence of depth cameras brings new opportunities for action and event recognition. In this talk, we will introduce the Kinect camera, which collects RGB+D (color + depth) image data and briefly overview its applications. Then we present two topics from our recent work at ADSC on recognition of action and events using Kinect cameras. First, we describe a homemonitoring oriented human activity recognition benchmark database. This database aims at encouraging research on human activity recognition based on RGB+D sensors. We present two RGB+D fusion schemes developed from two state-of-the-art feature representation methods for action recognition. Experimental results show superior accuracies achieved by using the depth feature. Second, we develop a RGB+D feature fusion method for video event detection, for the application of fall prevention in hospital ward. A multi-channel, multi-modal and multi-kernel based fusion scheme is proposed for event detection. Experimental results demonstrate the high accuracy and low false-alarm rate achieved in event detection.