| Video target tracking is one of the most urgent issues that need to be solved in computer vision field, it is the foundation of a series of following work. Target tracking has been researched for decades, and many efficient tracking algorithms have been proposed. However these algorithms can not be applied in many ordinary situations, because the target need to be tracked is usually interfered by many factors, such as illumination variation, fast motion, motion blur, in-plane rotation, out-of-plane rotation, occlusion and so on. It is difficult to propose an efficient, robust algorithm which can be applied to any situations. In this paper we did some thorough research on those algorithms that combine with the local context surrounding of the target.The spatio-temporal context(STC) was proposed by Zhang et al.[27] in 2014. It was based on the generative appearance model, and integrated with tracking, learning and detection. In STC, Zhang added background information which can be spatial location in generative appearance model, and adopt Fast Fourier Transform(FFT) to reduce calculation, so this algorithm could achieve real-time, robust and efficient tracking result. But STC still has some weaknesses:(1)It couldn't resist strong interference. If the target suffers strong interference in the process of tracking, the tracking result would appear drift;(2)It couldn't find target again in some extreme cases. The reason was that STC was a local search algorithm, it is impossible to find the target when the location where the target reappears oversteps the local context which was used to search the center of the target;(3)Drift problem often occurred on appearance model. The algorithm couldn't realize absolute accuracy on tracking especially when suffering strong interference. STC was based on Markov temporal smoothness assumption which would learn all tracking result of frames indiscriminately. So some background features have added into appearance model. As the appearance model updating persistently, drift prblems would occur and tracking fails.Aiming to solve the above problems, we did a series of research. Our research contents and results are as follows:(1) Propose an improved STC algorithm which is combined with particle filter. The new algorithm adds strong interference detection and particle filter to STC, based on the detailed analysis of STC algorithm. When the algorithm detect strong interference during the process of tracking, it would use particle filter to correct the predictive estimation result of STC. It effectively enhance the tracking robustness.(2) Propose an improved STC algorithm which adapts adaptive structure model. This paper presents a new graphical model named adaptive structure model which can be applied to online algorithm. First it should save a set number of snapshot templates, then it would use all of snapshot templates and a normal template together to estimate the result when a new frame comes. When one of snapshot templates is better than normal template, it would use the snapshot template to replace the normal template. So it can effectively restrain the error accumulation.(3) This paper uses the two improved STC algorithms to do experiments on Tracker Benchmark v1.0[1] dataset. The experiments show that, compared with STC, the average accuracy is improved from 37.88% to 46.39% and 42.02% respectively. The average center location error is decreased from 85.53 to 49.61 and 62.78 respectively. The average frame rate decreased from 45.89 fps to 44.91 fps and 36.64 fps respectively. Both of our algorithms meet the requirements of real-time tracking. In conclusion, our two improved algorithms in this paper effectively enhance the tracking robustness. |
PDF Full Text Request