Research On Target Tracking Algorithms Based On Deformable Model And Target Feature

With the development of Computer vision and image processing technology, Video-tracking technology has been more and more widely used in the fields of Engineering of Security and Protection System, Precision-Guided Munitions, and National Defense etc, at the same time a research highlight in the fields of automatic control, pattern recognition and computer vision. This thesis mainly conducts researches on algorithms about target detecting and target tracking under the background of video surveillance.First of all, the mainstream algorithms in fields of target detecting and tracking are deeply analysised in this thesis. Then, an improved target detecting algorithm well suited to tracking system is proposed. It combines accumulative frame-to-frame differences and background updating strategy, which can effectively overcome weakness of inaccuracy of extracting background and real-time problem of background updating. Thus, this algorithm has a certain capability of interference rejection when there is interference target.Additionally, this thesis researches on two mainstream target tracking algorithms. The first one is particle filter target algorithm based on GVF-Snake and the other is multiple kernels tracking algorithm based on histogram of gradient direction. The first mentioned is Particle Filter algorithm based on GVF-Snake. Comparing with traditional Snake, the GVF-Snake takes the advantage of powerful searching ability and it is not sensitive to initial points. The true contours of the objects can be converged well by means of GVF-Snake. Meanwhile, in consideration of complexity of tracking scene, the algorithm adjusts number of snake points in order to track moving and deformable object adaptively. In the process of Particle Filter, this thesis chooses the contour of target as target feature, and research on tracking of deformable target. The tracking experiment proves that this algorithm could track deformable target to a certain extend. Multiple kernels tracking algorithm based on histogram of gradient direction introduces the advantage of histogram of gradient direction in target detecting into the field of target tracking. On account of target blocked partly, the algorithm divides the target region into blocks and extracts kernel weighted histograms of oriented gradients for each block. The similarity between target model and candidate model is measured by the sum of Bhattacharyya coefficients of all the corresponding histograms. Then by maximizing the similarity using the Mean Shift tracking algorithm, the object is tracked. Also this method can make sure that part information of target will not be dropped by the way of reducing weight of edge pixel. Experiments on the tracking have proved that this algorithm above mentioned can achieve target tracking, have more accurate location of object and have a certain robustness to block target.Finally, this thesis integrats image board and develops auto-detecting and tracking system which relies on camera and PTZ devices. Also, the two tracking algorithms proposed in this thesis are validated on the platform for the practicality utility.