Research On Object Tracking Algorithm Based On Multiple Instance Learning

Visual object tracking is a very important research direction in computer vision field, which combines the advanced technology and research results of the pattern recognition, artificial intelligence and image processing. Visual object tracking has a very important practical value in many areas, at present it has already widely applied to the military, weapon system, traffic control, intelligent vehicles and safety monitoring and so on. However, it is still a challenging problems to track a target in real world environment because there are many influence factors such as occlusion,illumination change, motion blur and clutter background. In this paper, to solve these problems, multiple instance learning was applied in object tracking algorithm. Then, further research and analysis on tracking algorithm were taken which were based on multiple instance learning, and two kinds of improved tracking algorithm were proposed. At last, the results of simulation show that the robustness and accuracy of the algorithm were improved.Firstly, three main components in object tracking algorithm were analyzed, online learning tracking algorithm and combined with a variety of tracking algorithms were research focuses, which was found by the research of commonly used object tracking algorithm.Secondly, online Ada Boost tracking algorithm was studied, which was based on online learning and ensemble learning algorithm. But the simulation experiments show that it is difficult to select samples in online Ada Boost tracking algorithm, then the further research on multiple instance learning shows that: multiple instance learning can solve the problem of samples selection.Thirdly, multiple instance learning was applied into tracking algorithm to overcome the difficulties in the selection samples. And considering the difficulty in setting searching area, the object movement information was modeled, then according to its ability of classification the weak classifiers were weighted to improve the classification ability of strong classifiers. But the accumulation of error is still existence which is caused by self-training.Finally, in order to avoid the accumulation of error caused by self-training, co-training algorithm was combined, classifiers in two redundant characteristic views were trained and updated, and instances in sampled bags were weighted by its importance, then the tracking algorithm based on online multiple instance learning and co-training was proposed. The simulation experiments show that the algorithm in this paper has been improved in accuracy and robustness.