Multiple camera-based codebooks for object detection under sudden illumination change

TL;DR: Experimental results show that the proposed foreground segmentation method gives better performance compared to the single camera-based counterparts and other conventional approaches, and is also robust to shadows.

Abstract: In many computer vision applications, identification of moving objects is a critical task. It involves classification of a pixel into either foreground or background. Background subtraction is a common approach used to achieve such classifications to remove background from the current frame. Background subtraction/modeling is extremely difficult due to illumination variations and the presence of shadow and/or occlusion. Single camera-based setup cannot perfectly handle all these problems. To overcome some of these problems, multiple camera-based backgrounds modeling system is proposed to extract multi-view objects. In this paper, homography and codebook-based approaches are utilized to detect the moving objects. Subsequently, a new heuristics is proposed which is quite robust to sudden lighting changes. The proposed method is also robust to shadows. Experimental results show that the proposed foreground segmentation method gives better performance compared to the single camera-based counterparts and other conventional approaches.