Real-time Gesture Recognition for Interactive Presentation with Depth Sensor

Jingjing Fu

Abstract: In this paper, we propose a method to enable real-time interaction between the projection contents and speaker through detecting and recognizing meaningful human gestures from depth maps captured by depth sensor, making projection screen as a kind of touch screen. Considering that depth noise and serious occlusion may ruin the construction of skeleton, our hand trajectory is derived from Potential Active Region. To cope with their inter-class and intra-class variations, hand trajectory is temporally segmented into movements, which are represented as Motion History Images. A novel set-based soft discriminative model is learned to recognize gestures from these movements. In addition, as it is a real-time system, a complexity reduction method is employed. The proposed approach is evaluated on our dataset and performs efficiently and robustly with 90% correct recognition rate.