VLSI Architecture for An Object Change Detector for Visual Sensors R. Aguilar-Ponce, J. Tessier, A. Baker, C. Emmela, J. Das, J. L. Tecpanecatl-Xihuitl, A. Kumar, and M. Bayoumi Center for Advance Computer Studies University of Louisiana at Lafayette PO Box 44330, Lafayette LA 70504-4330 USA ruth@louisiana.edu Abstract— Object detection is a crucial step in visual surveillance. Traditionally, object detection has been performed purely in software in surveillance systems. The problem of object detection, however, becomes critical in the upcoming wireless visual sensors because of size and power constraints. The need for low-power, small size, hardware implementations is greatly felt. This paper introduces a VLSI architecture for Wronskian Change Detector (WCD). Object detection is done through background subtraction. WCD offers regularity, low complexity and accuracy as well as global illumination changes independency. WCD can be employed in automated visual surveillance on buildings and adjacent parking lots. WCD replaces each pixel by a vector containing luminance value of the pixel and its surrounding area. A linear dependency test is applied to each vector to determine if a change has occurred. WCD is mapped into a 12 – Processing Element array with a fixed window value of 3×3. Design of each processing element is discussed in detail. Based on extensive search, no VLSI implementation of WCD has been reported previously I. INTRODUCTION There is an increasing demand for surveillance system in today’s daily life. From the technological-solution perspective, video surveillance has been widely employed for this purpose. Wireless visual sensors promise significant possibilities of performing surveillance at low cost and high speed. The problems of the traditional visual surveillance [1] are further exacerbated by the need to perform low-cost, low- power, and high-speed operations in sensors. These technical challenges include system design and configuration, architecture design, object detection, object identification, tracking and analysis, restrictions on network bandwidth, physical placement of cameras, installation cost, privacy concerns, and robustness to change of weather and lighting conditions. In this work, we focus on the object detection. Change detection plays a key role in real-time image analysis. Detection on the scene under observation includes moving objects, addition or removal of objects. Therefore, changes due to the change in illumination as well as noise must be disregarded. One key issue is robustness against illumination changes. Several approaches have been proposed over the years [2]. The most instinctive technique is frame differencing followed by thresholding. Change is detected if the difference of the corresponding pixels exceeds a preset threshold. The advantage of this technique is its low computational complexity, however it is very susceptible to noise and illumination changes. Median filter is one of the most popular background subtraction techniques [3]. Median of each pixel of all the frames in the buffer constitutes the background estimation. Background pixels are considered to be those that stay on more than half of the frames on the buffer. However, this technique requires a buffer large enough to store L frames. Recursive background techniques do not require a buffer of previous frames. In its place, they recursively update the background model based on each input frame. Any error in the background estimation can remain for a long period due to its recursive nature. The most popular recursive technique is Mixture of Gaussian (MoG) [4]. This method models each background pixel by a mixture of K Gaussian distributions (K is a number between 3 and 5). Different Gaussians are assumed to represent different colors. The weight parameter of the mixture represents the time proportions that those colors stay in the scene. The probable background colors are the ones that stay longer and more static. However, the technique is computationally intensive; its parameters require careful tuning and it is very sensitive to sudden changes in global illumination. Wronskian Change Detector employs the Wronskian of intensity ratios as a measure of change [5]. A large mean or large variance of the intensity ratios increases the Wronskian value. This method can detect object interiors and structural changes. Also, WCD is robust against illumination changes. WCD is a suitable algorithm to be implemented in real-time due to its low complexity. Also, this technique requires only one previous frame; therefore it is appropriate for applications where resources are limited. A comparison of the discussed methods is presented in Table I. WCD offers a tradeoff between complexity and 0-7803-9333-3/05/$20.00 ©2005 IEEE SIPS 2005 290 Authorized licensed use limited to: University of Louisiana at Lafayette. Downloaded on December 2, 2009 at 12:20 from IEEE Xplore. Restrictions apply.