Research On Online Video Object Tracking Algorithm In Presence Of Its Zoom And Occlusions

Target tracking in video, as a key problem in computer vision, has a wide range ofapplications in all aspects of people’s lives. There are many influencing factors forvideo target tracking such as lighting,occlusion, target motion or shape change, etc. sodeveloping an efficient algorithm for a robust tracking is a difficult and challenging task.This paper mainly focuses on researching one kind of tracking algorithms based ontracking-learning-tracking model.The Compressive Tracking (CT) algorithm proposed by Zhang K H etc in2012integrates tasks: tracking, studying and judging, which can achieve real-time andaccurate tracking results. The optimized multi-feature CT algorithm based on CTalgorithm presented by Zhu Q P etc in2013remedy its shortage of unbalanced featureselection for texture and gray feature, making the tracking results more robust. Althoughthe most successful of both the two methods are simple and efficient, they also havesome disadvantages:①the tracking window is fixed, so when the target size changesamong tracking, the probability distribution of the target will easily be influenced,doping too much background information or falling into local of the target, resulting inthe accumulation of tracking error;②without correctness judgment of the trackingresults,directly make the result be the basis for the follow-up video tracking. Oncedeviation happens, it will lead to growing tracking error,especially when the target isoccluded.According to the existing problems in multi-feature CT algorithm that the trackingresult will be affected seriously when the target size changes or the target is occluded,making a series of further research, the research contents and results are as follows:①Proposes an improved multi-feature CT algorithm in which the size of trackingwindow can adaptively changes as the target size changes, based on the analysis of theprinciple of multi-feature CT algorithm. The new method changes the collection methodof candidate targets sets from one scale into multiple scales, and changes the featureextraction method and feature representation method. Determine the tracking result outof the multi-scale candidate target sets by Bayes classifier according to the Gaussianprobability distribution of the features.It makes the scale of the tracking windowchanges correspondingly as the target scale changes, and resulting in a better trackingperformance. ②Proposes an improved solution based on multi-feature CT algorithm in order toreduce the bad influence on tracking caused by occlusion. After analyzing of the reasonwhy occlusion will result in tracking lost, this new method formulates the occlusiondetection mechanism and Kalman Filter prediction mechanism. The target is dividedinto four sub regions and each sub-region respectively trackes with multi-feature CTalgorithm and respectively updates the Gaussian probability distribution of the features.The tracking process for each frame produces two results: a) the back calculatedposition by the sub-region with the highest confidence by Bayes classification; b)thepredicted position by Kalman Filter. Then judge that weather the sub-region withhighest confidence is occluded. If so, take a) as the tracking result,else take the other.In addition, in order to reduce the impact by occlusion object, only the sub-regionswithout occlusion update the probability distribution of the features. The new methodeffectively reduces the adversely affection outcomes by occlusion problem.③The two improved algorithms based on multi-feature CT algorithm which arementioned above are verified and analyzed by experiments. The experiment results onvariety challenging video sequences compared with the CT algorithm and multi-featureCT algorithm shows the availability of the two improved methods in terms ofefficiency,accuracy and robustness.