Pattern Recoonition, Vol. 25, No. 10, pp. 1171 1180, 1992 Printed in Great Britain 0031 3203/92 $5.00+.00 Pergamon Press ktd © 1992 Pattern Recognition Society A VIDEO TRACKING SYSTEM WITH ADAPTIVE PREDICTORS LIANG-HWA CHEN and SHYANGCHANG+ Department of Electrical Engineering, National Tsing Hua University, Hsinchu, Taiwan, R.O.C. (Received 3 September 1991; in revised .[brm 3 February 1992; receited fi~r publication 14 February 1992) Abstract--A novel video tracking system with adaptive predictors is developed. In order to locate the moving target efficiently, a Hadamard transform locator is proposed. The accuracy of the target location is usually influenced by noise,hence an effectivemorphologicalnoise filter is invoked to removeit. Moreover, the recursive least square adaptive predictor (RLSAP) is adopted to predict robustly the future location of the target. After successful integration of the aforementioned subsystems, the moving target is always kept in the center of the observing window. Experimental results are given to demonstrate the effectiveness of this proposed system. Moving target tracking Azimuth and elevation Mathematical morphology Hadamard transform RLS adaptive predictor I. INTRODUCTION Moving target tracking is an interesting and useful research area in computer vision. It can be used in mili- tary and industrial applications. There have been many works concerning the moving target tracking problem. Gilbert et al. ~1) presented a real-time video tracking system which used extensive parallel processing to enable tracking of missiles in 512 × 512 pixel scenes at a rate of 60 frames s-1. The main components of this system consist of a Bayesian classifier, projection processor, finite state automaton and combination of linear and quadratic predictors. In order to implement these four components, specific bit-slice microproces- sors are adopted for parallel processing. Schalkoff and McVey t2) presented a video tracking algorithm based on two-dimensional (2D) affine transform and Taylor series expansion. Kabuka et al. ~3> presented another video tracking system based on optimizing control. Their algorithms required that the target be centered in the field of view (FOV) of the camera at the outset. Lee et al. ~+~ proposed a system for tracking a slowly moving ship in an infra-red image sequence based on a multi-frame detection concept. Since there is no tracking motion of the camera involved, it is not in the same category of the aforementioned tracking systems. There were some other algorithms which were based on statistical techniques. Milstein and Lazicky~5'6~ used a maximum likelihood method to detect and track the target in 20 x 20 pixel subwindows in a 100 × 100 pixel FOV. Mohanty <vJ also tracked low intensity point targets in space based on the maximum likelihood ratio. Winkler and Vattrodt 18~ used a conspicuousness detector in an interactive manner to track the target t Author to whom all correspondenceshould be addressed. with a local maximum conspicuousness in a tracking window. One major drawback of these statistical approaches is that a priori information about the statistical distribution of the target and background are needed. Moreover, the amount of computation is usually very heavy. The correlation matching technique was also applied to track algorithms. McVey and Woodlard 191 used cross correlation to derive TV- camera control signals for target tracking. Dovorny- chenko and Mack I1°) used correlation measures with Lp norms to track partially obscured targets. However, the problem with matching via correlation is that it is very time consuming. Another major drawback with all the conventional approaches is that the motion parameters of the target are defined on the image plane. This coordinate system is actually time varying. It varies with the pointing direction of the camera. Thus, the coordinates of the target calculated at different observation times are based on different coordinate systems. As a result, we cannot use the information of previous coordinates to adaptively predict the future location of the target. In this paper, a video tracking system with adaptive predictors is proposed. The motion parameters of the target will be transformed to the world coordinates from its motion on the image plane. Hence, the motion parameters can be estimated via adaptive predictors using the information in the world coordinates. In addition, the a priori statistical information about the target and its background are no longer required. Furthermore, the target image does not need to be centered in the FOV at the outset. The block diagram of the proposed system is described in Fig. 1. First, consecutive images are acquired while the camera's direction is kept fixed. From these consecutive images, the moving target's image will be detected if there is one. Since registration noise will definitely occur, it ll71