Multi scale feature point tracking

Feature point tracking is one of the most important subjects in machine vision due to its abundant applications. The advantage of this sort of tracking compared with other tracking methods is in extracting exact position information from tracking object´s components. Three important problems in this issue are accuracy, speed and stability of the tracking system subject to illumination variations, noise and large displacements. This paper tends to present a new algorithm for feature point tracking with sub-pixel accuracy. To obtain proper stability and accuracy, multi-scale feature point description is used. This description considers both exact location of the feature point and its neighborhood simultaneously to obtain high accuracy together with high stability. Also, to reach proper speed, tree search algorithm is used. This type of searching prevents local optimums because of searching the probable location for the feature point, completely. In addition, using the hierarchical searching method, results in faster tracking. The proposed method is compared with standard KLT and pyramid KLT on real videos to be evaluated in quality and speed.