Object Tracking Algorithm Based On Correlation Filters

In recent years, due to the application of target tracking technology in intelligent transportation systems, human-computer interaction and so on, the research of target tracking technology has made great progress, and a lot of excellent tracking algorithms have emerged, but most of the target tracking algorithms are still faced with a series of challenges such as target scale transformation, target occlusion and illumination changes and other factors which often leads to target tracking shift or even failure. In order to solve the problem of target occlusion and scale transformation in object tracking, this paper makes a deep research on the problem of feature fusion and adaptive scale transformation, the main results are as follows:(1) An adaptive weighted fusion method based on color feature and HOG feature is proposed, then, according to the magnitude of the response values of the two kinds of feature, the weights of the two kinds of feature distribution are adaptively assigned. This method can solve the problem of poor robustness of single feature.(2) The effect of target appearance mode land learning rate of classifier parameters on target tracking is analyzed. By using the frame difference method to determine the moving speed of the target, a method of determining the learning rate according to the different movement speed of the target is designed. This method can solve the problem of object occlusion well.(3) A scale prediction method is given according to the target location of feature fusion, the HOG features of samples with different scale sizes are extracted, By the means of training least squares method classifier of kernel function, the maximum value of the scale output response is obtained and the prediction of the target scale is completed. The method can change the size of the rectangle tracking frame according to the size of the target, thus improving the tracking accuracy.(4) The above three kinds of algorithm together to form the final algorithm in this paper, In the experiment,9 groups of challenging standard video sequences were tested. The experimental results show that the mean center position error (CLE) of the proposed algorithm is 5.18 pixels, the average distance accuracy (DP) was 95.09%, the average accuracy of overlap (OP) up to 96.44%. Compared with the best among existing 4 tracking methods based on correlation filter, The average position error in this method (CLE) was reduced by 4.09 pixels, overlapping accuracy (OP) increased by 22.24%, the average distance accuracy (DP) increased by 6.75%; Under the complex conditions such as object occlusion and scale transformation, the algorithm is still able to track the moving target stably and accurately, which has very important theoretical value and applied research value.