Visual Object Tracking Based On Kernerl Correlation Filter And Visual Attention With Multi-Feature Dynamic Integration

Visual object tracking technology as an important research direction of machine vision is widely used in intelligent monitoring, human behavior analysis, intelligent robot and other fields. And the existing tracking method has great room to improve on the speed and accuracy for the complex scenarios and uncertainties such as motion blur, morphological changes,object occlusion, background interference and so on. Therefore, it is still a valuable task to research a new tracking algorithm with real-time performance and strong overall tracking ability.Based on the predecessors' work, this paper mainly studies on three aspects including feature integration, dynamic allocation of feature weight, and human visual habit, and puts forward some improvement strategies, which are summarized as follows:1. In order to improve the adaptability of the visual object tracking algorithm, a multi?feature integration algorithm based on KCF is constructed by integrating the idea of SAMF,DSST and Staple algorithm ofter deeply analyzing on advantages and disadvantages of the KCF algorithm. Which is based on KCF by coalescing the three local features of gray, HOG and CN, combining the color histogram tracking strategy in Staple, and using the method of DSST to realizing the scale estimation. And the overall tracking ability is improved by combining the advantages of multiple features and various algorithm.2. In order to further exploit the advantages of multi-feature integration algorithm,especially the advantages of tracking method by color histogram for motion blur and morphological changes, a new method of adaptive dynamic assignment weighting coefficient of color histogram is proposed. In which several tracking results are obtained by setting different values of weighting coefficients. Then, the obtained results are scored by the quadratic KCF without the background. Fnally, the weight coefficient of the highest score is selected. The method can adaptively select a better weight coefficient in different scenarios,and further improves the adaptability of the algorithm.3. Inspired by the human visual mechanism, the method of calculating the probability of the pixel belonging to the object in the color histogram tracking is improved, which increases emphasis on significant color of the object, and further enhances the overall tracking performance of the algorithm, especially has a very good performance in the scene of severe deformation.Finally, the algorithm proposed in the paper can get 5.1% improvement in median distance precision and 3.7% improvement in median overlap precision compared with Staple(2016 CVPR) algorithm in experiments on OTB100 and VOT2016 dataset with a total of 160 videos, and can achieve real-time tracking.