Combining ARF and OR-PCA for robust background subtraction of noisy videos

Background subtraction is a fundamental pre-processing step for many computer vision applications. In addition to cope with dynamic background scenes, bad weather conditions such as rainy or snowy environments and global illumination conditions such as light switch on/off are still major challenging problems. Traditional state of the art methods, such as Robust Principal Component Analysis fail to deliver promising results under these worst conditions. This is due to the lack of global preprocessing or post-processing steps, incorrect low-dimensional subspace basis called low-rank matrix estimation, and memory or computational complexities for processing high dimensional data and hence the system does not perform an accurate foreground segmentation. To handle these challenges, this paper presents an input video denoising strategy to cope noisy videos in rainy or snowy conditions. A real time Active Random Field constraint is exploited using probabilistic spatial neighborhood system for image denoising. After that, Online Robust Principal Component Analysis is used to separate the low-rank and sparse component from denoised frames. In addition, a color transfer function is employed between the low-rank and the denoised image for handling abruptly changing lighting conditions, which is a very useful technique for surveillance agents to handle the night time videos. Experimental evaluations, under bad weather conditions using two challenging datasets such as I-LIDS and Change Detection 2014, demonstrate the effectiveness of the proposed method as compared to the existing approaches.