On-line Semi-supervised Boosting With Scale Adaptive For Robust Object Tracking

Object tracking is a hot issue in the field of computer vision, it has a wide range of applications in surveillance, human computer interaction and medical imaging.Although object tracking has been studied for several decades, and much progress has been made in recent years. But in the influence of illumination variation, occlusion, as well as background clutters, it is still a hard problem under complex dynamic scenes.Due to good adaptability, object tracking algorithms based on online learning achieve success in the field of object tracking application, it has become a mainstream technology in the field, among them, the most famous work is on-line boosting tracking algorithm. however, the main defects of the algorithm are the object "drift" during the tracking process and how to adapt to the changes of object size, in order to solve above problems, based on on-line boosting tracking model, this article introduce semi-supervised learning strategies and scale adaptive operator to improve the performance of tracking algorithm, the main works are as follows:(1) By introducing a semi-supervised learning strategies, an on-line semi-supervised boosting tracking method was introduced in detail the steps and advantages in solving the problem of "drift". Secondly, in order to solve the problem of scale variation of object during the process of tracking, two robust target tracking algorithms with adaptive window are proposed.(2) A visual image information based adaptive tracking window adjustment method for object tracking is proposed. Because the image information under different scale spaces will change with the variation of the scale, the key feature points such as edges and corners, the maximum points and minimum points can be used to evaluate the amount of information in specific area so it will be favor to realize the adaptive adjustment of the tracking window. At the same time by on-line boosting with a semi-supervised learning strategy, the object drift problem will be solved.(3) An on-line robust object tracking with scale adaption based on weighted image is introduced. By analyzing the moments of the weight image that based on statistical features of object grayscale or color histogram features, the adaptive of window can be achieved. At the same time, using labeled training samples a predefined classifierpH is structured, moreover N individual classifiers, 1,...,nH n ?N are attained with unlabeled samples, the proposed trackingalgorithm not only inherits the advantages of online object tracking algorithm for variable lighting, background and appearance but also shows good robustness.