International Journal of Computer Vision and Signal Processing, 4(1), 22-28(2014) ORIGINAL ARTICLE Extraction of Potential Sunny Region for Background Subtraction under Sudden Illumination Changes Ikuhisa Mitsugami * The Institute of Scientific and Industrial Research Osaka University, Japan Hiromasa Fukui, Michihiko Minoh Academic Center for Computing and Media Studies Kyoto University, Japan ISSN: 2186-1390 (Online) http://www.ijcvsp.com Abstract This paper proposes a novel background subtraction method robust for sudden illumination changes that often happen in outdoor scenes. The method first estimates regions where are sunny regions or would become sunny regions when the sun is not behind clouds, which we call “po- tential sunny regions.” For the estimation, spatio-temporal analysis is applied to image sequences of the recent days of a target day consid- ering the periodicity of the sun’s movement. Once the potential sunny regions are obtained, they are used for judging if the sudden illumination change happens in the target scene. When it happens, then the illumi- nation changes within the sunny regions are suppressed to obtain better subtraction results. Experimental results in several outdoor scenes show effectiveness of the proposed method. Keywords: Background subtraction, long-term observation, spatio-temporal analysis c  2014, IJCVSP, CNSER. All Rights Reserved Article History: Received: 26 August 2014 Revised: 18 December 2014 Accepted: 19 December 2014 Published Online: 20 December 2014 1. Introduction Background subtraction is one of the fundamental tech- niques for many scene understanding application, espe- cially for video surveillance. It is useful because it can provide foreground segmentation without any prior about the foreground, and so there are still so many studies about the background subtraction [1]. Among the studies, pixel-wise background modeling is the most popular approach. In this approach, distribution of pixel value is modeled by such as an average or median of the recent pixel value history [2, 3, 4], Gaussian mixture model (GMM) [5], kernel density estimation [6], Parzen density estimation [7], and so on. Liu et al. proposes an- other model named Effect Components Description (ECD) [8]. * Corresponding author Email addresses: mitsugami@am.sanken.osaka-u.ac.jp (Ikuhisa Mitsugami), fukui@mm.media.kyoto-u.ac.jp (Hiromasa Fukui), minoh@media.kyoto-u.ac.jp (Michihiko Minoh) These pixel-wise methods have an advantage that the sub- traction results do not loose resolution of their input im- ages; they output the results with the same resolution as their input. On the other hand, however, they have an essential problem that they cannot discriminate be- tween pixel value changes by foregrounds and illumination changes, since each model knows only history of the corre- sponding pixel. Indeed such illumination change is learnt in the model after a certain period, but it is theoretically impossible to adapt to the change immediately. To over- come the problem, other studies adopt region-based back- ground modeling [9, 10, 11]. In these methods, not each pixel but relationship among neighboring pixels is modeled, and they are more robust for the illumination change. The resolution of their output is, however, usually decreased compared with their input. In this paper, therefore, we proposed a novel back- ground subtraction method that solve this trade-off; the proposed method gives as high resolution output images 22