Region-of-importance detection based on fusion of audio and video
Tao Wu, Cuong T. Vu, Qi Cheng, Damon M. Chandler · 2009
In this paper, a new framework is proposed for autonomous universal surveillance based on video and audio data. ¿Universal¿ indicates no specification of targets of interest. Instead, regions of importance (ROIs) in a scene should be detected. Specifically, in the video domain, a frame-based main subject detection is proposed based on adaptive selection of low-level features. In the audio domain, a time-delay-based direction of arrival estimation scheme is adopted. The outputs of video and audio processing are fused in a probabilistic framework to generate more refined ROIs. Experimental results demonstrate the effectiveness of the proposed scheme in ROI detection.