REAL-TIME OBJECT DETECTION USING TWO BACKGROUND MODELS UNDER SHAKING CAMERA

Abstract

In this paper, we propose a novel approach to detect moving objects by two background models, multiple background model (MBM) and temporal median background (TMB), from hand-taken image sequence. For this purpose, we record image sequences by hand-held camera without tripod so every frame has variation between consecutive frames. A pixel-based background model is fragile while image sequence has variation. Therefore we calculate the camera movement using correlation between two consecutive images and it helps us to generate MBM under shaking camera. The computational cost of correlation quickly increases if image resolution increases. Hence, we use edge segments to reduce computational cost. These edge segments are gathered by Sobel operator and those are distinctive spatial features to calculate similarity between two regions, belonging to current and previous images, organized by neighbors of edge segments. Based on the similarity result, we obtain a set of best matched regions, centroids of matched regions, and displacement vectors from each pair of previous and current images. Each displacement vector in a set describes the transition of each matched region in the image pair. Using the highest density of displacement vector histogram, we choose the camera motion vector, indicates camera movement between consecutive frames. According to the camera motion vector, every pixel in a current image is related to different position pixels in a previous image. The pixel relation is used to generate MBM in this paper, unlike original MBM [Xiao, M., Han, C. and Kang, K. [2006]. Proc. Int. Conf. Information Fuscon, pp. 1–7.]. The MBM algorithm classifies the variation of pixel values in frame sequence to several clusters. Classification of varying pixel values to several clusters is similar with mixture of gaussian (MOG). Nevertheless, MBM has low cost to calculate because it does not need to estimate parameter. However, MBM is not sensitive to short period changes. Therefore, we use TMB to support MBM. The experimental result shows that proposed algorithm successfully detects moving objects using background subtraction less than 25 ms per frame when camera has 2D translation.