Robust dual-model object tracking with camera in motion

This paper proposes a robust dual-model matching and model update algorithm for object tracking with camera in motion. With the proposed technique, a short-term model is matched then updated each frame when it is matched to reflect the most recent changes of the tracked object in illumination, size and deformation. Furthermore, a long-term model is maintained in a period of time then updated in a lower frequency to against the interference of rapid and temporal change of the above factors. In addition, a statistic-based sub-region update strategy rather than global update for both short-term and longterm tracking models is utilized. The proposed algorithm is intensity-based, but it can be extended to any feature matching, such as color, texture and shape appearance. An embedded version of this algorithm has been implemented with Texas Instruments' DM6446-based DSP system. Extensive experiments and practical applications in different situations confirm the robustness and reliability of the proposed method.