Fast algorithms for video target detection and tracking

This paper addresses two problems commonly associated with video target tracking system. First, video target detection and tracking usually require extensive searching in a large space to find the best matches for preregistered templates. Existing fast search methods cannot guarantee a global optimal match, which results in substandard performance. To obtain a true global match, a full search at the pixel or sub-pixel level is required. Obviously, this introduces significant computational overhead, which limits the implementation of these algorithms in real-time applications. In this paper, we propose a fast method to compute two-dimensional normalized cross-correlations to efficiently find the global optimal match result from a large image area. Comparisons and complexity analysis are provided to show the efficiency of the proposed algorithm. Second, another challenge commonly faced by detection and tracking systems is the accurate detection of target orientation in a twodimensional image. This problem is motivated by applications where the walk-in and walk-out people need to be detected and a fast image registration method is needed to compensate the change in rotation, translation and size, which is natural since the target's distance from the camera is changing dramatically. To address this issue, we propose a novel and efficient eigenvector-based method to detect target orientation and apply it into automatic human recognition system. Experimental and real-world test results verify that the proposed fast algorithm achieves similar accuracy as the recursive registration method which is computationally expensive.