Visual Tracking Algorithm Based On Template Clustering And Superpixel Discriminant

Target appearance representation is a primary and key issues in visual tracking. Considering the representation of target appearance, we establish a stable target ap-pearance representation which could adapt the target changing. In order to suppress-ing the background interference, background representation of the search area around target was built to describe the complex background information. First, candidates were extracted from the search area in the tracking frame. Second, a likelihood function then would be constructed by the distance between candidates with positive class and candidates with negative class. At last, we treat the candidate with max likelihood function value as the true target in tracking.The template of most tracking algorithms is difficult to adapt the appearance changes of non-rigid objects due to high redundancy. To overcome this problem, a tracking method with online templates clustering based on particles filter is proposed. First of all, the positive and negative template sets are established to describe the target and the background. Secondly, candidates are extracted based on the dynamic model. And then the likelihood function is built through the within-class distance between candidate and template sets and the between-class distance between positive sets and negative sets. Finally, the best candidate is considered as the tracking result according to the maximum a posteriori probability (MAP). The main idea of online template clustering is as follows:First, the cluster radius is determined by the state class which is produced by a series of continuous target states in a certain range. Second, the positive template set combined with the recent several tracking results are used for clustering by using the mean shift iterative method and the above cluster radius. Last, the updated positive template set consists of the new cluster centers. In complex situations, experiments show that this method can retain different appear-ance states of target and track the target accurately.The global polluted template introduced errors in tracking algorithm and will cause a tracking drift problem. A tracking algorithm using discriminant superpixel attribute was proposed. Firstly, the image was divided into many superpixels. Sec-ondly the over-segmentation superpixels will be merged into less superpixels. The merged superpixels confidence could be calculate with the distance of superpixel to the target's centre. Thirdly a decision tree was trained with superpixels and its' labels of target and background. Then we would obtain an image confidence map with su-perpixels discriminant by the decision tree classifier. Finally many candidate parti-cles were sampled in the confidence map and select the best candidate particle as the true target. Experiments show that this method could deal with target occlusion and track the target accurately.