Several algorithms have been developed for the tracking of a moving target in image sequences. In these tracking algorithms, a centroid tracking algorithm and a correlation tracker are the most popular ones. The centroid tracker determines a target aim point by computing the geometric or intensity centroid of the target object based on the target segmentation method. In the correlation tracker, the motion of a block of pixels, termed a reference block, is estimated by looking for the most similar block of pixels in the subsequent frames. This dissertation studies on new automatic tracking algorithms of moving targets in image sequences, which include an intelligent centroid tracking algorithm and a robust correlation tracker.
The performance of the centroid tracker depends on the following factors: (1) efficient real-time preprocessing technique, (2) exact segmentation algorithm, and (3) intelligent control of a tracking window size, etc.. Previous segmentation algorithms for the centroid tracker utilize only an intensity feature to segment a moving target from background images. And, they assume that all the related probability density functions of the target and the background are Gaussian ones in most cases. Ordinary segmentation algorithms often produce unstable results in real unconstrained outdoor scenes. In real tracking environments, the assumed Gaussian pdf`s are very different from those of real image sequences. Thus, it may be almost impossible to extract the target exactly from the cluttered image sequences.
This dissertation proposes an efficient real-time preprocessing method in order to enhance the distinction between the objects of interest and their local backgrounds. In addition, a real-time adaptive segmentation method based on new distance features is also proposed for the intelligent centroid tracker. The novel features include spatial distances between the predicted center pixel of a target from a tracking filter and each pixel for the extraction of a moving target. Then, the proposed distance features restrict clutters with target-like intensity from entering the tracking window, and have low computational complexity for real-time applications compared with other complex feature-based methods.
Furthermore, a fuzzy decision classifier which minimizes a fuzzy probability of a pixel classification error is also suggested to utilize the merits of the fuzzy set theory. Effective fuzzy rules and fuzzy membership functions are devised for the determination of the fuzzy decision classifier. Additionally, an intelligent control of the tracking window size is included to remove similar intensity clutters in adjacent to the moving target.
A correlation tracking algorithm seeks to align the incoming target image with a reference block image of target, but has critical problems, called false peaks and drift phenomenon (in other words correlator walk-off). The false peaks are generally caused by highly correlated background pixels with similar intensity of a moving target and the drift phenomenon occurres due to the accumulation of small tracking errors from frame to frame in the correlation process. A problem of the conventional correlation measures in the correlation process is that all the pixels of a reference block image are equally treated in the computation of the correlation measures, irrespective of a target or background pixel. Therefore, the more the reference block image includes background pixels, the more probability of false peaks is increased due to the cross-correlation between the background pixels of the reference block and the input image in the search area.
In this dissertation, a novel correlation measure with selective attentional property is proposed for the robust correlation tracker. The selective attentional correlation measure has different considerations according to the target and background pixels in the matching process, which conform with the selective attentional property of human visual system. A simple mathematical analysis is provided and it shows better performance compared with those of conventional measures. A neuro-similarity measure network of the block matching algorithm is also proposed
for the precision target tracking. The adaptive weight updating algorithm compensates for the changes of the reference image on a frame-by-frame basis to obtain an updated reference image. In addition, a robust shape matching tracker with gradient preprocessor combined by a drift removal compensator is also suggested to overcome the walk-off problem. The drift compensator adaptively controls the template size according to the target size of interest.
Finally, in order to overcome the limitations inherent in any single-mode tracker and to combine the proposed trackers (the intelligent centroid tracker and the robust correlation tracker), this dissertation addresses a multimode tracking system and image metrics which are qualitative measures that characterize the quality and the content of image or the scene. Throughout experiments of various synthetic and real images, the intelligence and robustness of the proposed tracking algorithms over the conventional tracking ones are demonstrated and discussed.