Selective object tracker: Evaluation and enhancement

摘要

The following paper introduces a slightly enhanced version of the Selective approach for tracking individuals in video sequences. In this version, tracking is achieved by evaluating candidate regions located in the eight adjacent neighbors in addition to the current position. Evaluation is handled using the Bhattacharyya distance. A process of motion estimation is incorporated to allow readjustment of the search region's size in order to handle false detections. To validate our approach, several experiments were conducted. Firstly, our new approach is compared to the Mean Shift algorithm on a personal dataset, and then secondly the algorithm is evaluated using a set of benchmark videos taken from publicly available datasets, handling different real-case scenarios.