A Moving Object Detection Algorithm for Smart Cameras

Abstract

In video surveillance systems, more intelligent functions are being embedded into cameras. In this paper, we propose a novel approach for detecting moving objects with fewer false alarms and low computational complexity. Contrary to traditional temporal differencing, we suggest a new feature based on signed difference, which represents the pattern of motion in a given region. The pattern is measured by calculating the Earth Mover's Distance between regions with opposite signs in the signed difference. The robustness of the perceptive metric is exploited to detect moving objects accurately. We also propose an efficient algorithm to approximate the value along four given directions and to locate the regions of moving objects. Experimental results show that, without prior knowledge or training, our method is able to detect moving objects with very low false alarm rate even under the challenging environments. In addition, the fast algorithm can be embedded into cameras and detect moving objects accurately in real time.