Block And Scale Adaptive Tracking Algorithm Based On Kernelized Correlation Filters

The visual target tracking technology is a research hot of computer vision,it is widely used in vision monitoring,human-computer interaction,intelligent transportation,military guidance and other fields.Many theories and tracking algorithms have been proposed,and the tracking accuracy and robustness greatly improved.However,in the actual tracking scene,how to gain the real-time performance and robustness of the algorithm is still a challenging problem because the target is affected by occlusion,scale change and other factors.For this reason,this paper proposes a block and scale adaptive tracking algorithm based on Kernelized Correlation Filters.The main work is as follows:1)Block fusion tracking algorithm based on Kernelized Correlation Filters(BKCF).Firstly,according to the size and width-height ratios of the target to adaptively divide it blocks,and extract the Histogram of Oriented Gradient and Color Name features for each block;Secondly,represent the each block with the fusion of Oriented Gradient and Color Name features,and use the kernel correlation filtering tracker to get the position with maximum response of each block;Thirdly,through the coordinate geometric relationship between the each block and the original target to calculate the coordinate of candidate target;Finally,the final target position is obtained by weighted average the coordinates of all coordinates.In addition,we take the Peak to Side lobe Ratio of a block's response curve as well as the distance between its response position and the final target position to judge its validity,and adaptively update the valid block.2)Scale adaptive tracking algorithm based on BKCF.Firstly,the final position of the target in frame t is tracked by the BKCF.Then take the sub-block target as the center to generate samples with multiple scales and extract the feature for the samples by Scale Invariant Feature Transform(SIFT).Construct a one-dimensional scale filter for each sub-block as the output of the scale pyramid,so as to obtain the optimal scale of the sub-block.Finally,according to the size of the maximum response value modeled by the SIFT of each sub-block,the weight is assigned to the optimal scale corresponding to the sub-block target,and the final target scale is obtained through the adaptive weighted average.Experimental results on many image sequence show that the proposed two algorithms can deal well with problems of target occlusion,Scale Variation,deformation,background clutters,and so on,and their performance is generally better than that of the compared methods.