A Motion Detection Method For Large Noise And Low Illumination Environment

The technology of motion detection is regarded as the foundation of high level video processing, such as tracking, object detection, behavior understanding and so on. The research and applications of motion detection are extensive and in-depth. However, the applicability and performance of motion detection methods are dramatically degraded with the increase in noise and decrease in light. Employing denoising methods before motion detection alleviates rather than solves this problem, particularly when an image or video is seriously polluted. In this paper, we propose a dictionary-based background subtraction approach that is robust against noise. The proposed method adopt the characteristic that noise projecting over the redundant dictionary is random and out of order to achieve the purpose on motion detecting under the above condition. This research includes the following aspects:1. Through analyzing sparse theory recently applied to the domain of motion detection, this research proposes the assumptions of motion detection under the condition of large noise and low illumination. Besides, the identity of noise over redundant dictionary is discussed in details.2. This paper proposes a background modeling method and casts the process of background modeling as a linear and sparse combination of atoms in a pre-learned dictionary with 1l norm. A strategy for atoms in redundant dictionary updating is proposed to guarantee the background model to agree with dynamic environment.3. A novel foreground detection method is implemented to compare the difference in sparse coefficients between the current frame and the background model. This research proposes a Structural Similarity Constraint to restrain the sensitivity of low illumination environment against the light. Then, the sliding window post-processing is applied to further optimize the detection results.4. Experimental results on synthetic noise, actual noise and dynamic background environment demonstrate the robustness of the proposed approach in terms of both qualitative and quantitative evaluations by comparing with classic, mainstream and other motion detection methods based on sparse theory.