LAUSR

Extracting stable features to enhance object representation has proved to be very effective in improving the performance of object tracking. To achieve this, mining techniques, such as $K$ -means clustering and data associating, are often adopted. However, $K$ -means clustering needs the pre-set number of clusters. Real scenarios (heavy occlusion and so on) often make the tracker lose the target object. To handle these problems, we propose an intraframe clustering and interframe association (ICIA)-based stable feature mining algorithm for object tracking. The value (in HSV space) peak contour is employed to automatically estimate the number of clusters and classify value and saturation colors of the object region to get connected subregions. Every subregion is described with observation and increment models. Multi-feature distances-based subregion association, between the current object template and the current observation, is then utilized to mine stable subregion pairs and obtain feature change ratio. Stable subregion displacements, and current detected and historical trajectories are systematically fused to locate the object. And, stable and unstable subregion features are updated separately to restrain the accumulative error. Experimental comparisons are conducted on six test sequences. Compared with several relevant state-of-the-art algorithms, the proposed ICIA tracker most accurately locates objects in four sequences and shows the second-best performance in the other two sequences with only less 1 pixel distance difference than the best method.