Robust visual tracking using incremental appearance descriptor update

Appearance descriptor is a fundamental component for visual tracking. But the targetpsilas appearance always changes over time in video stream due to variations in illumination, pose, scale and so on. In this paper, we combine particle filter with incremental covariance descriptor update for robust visual tracking. We employ a weight factor update mechanism to account for the contributions of previous observations. The weight factor is computed as the likelihood between observations and reference covariance descriptor. We update the covariance descriptor dynamically during the whole tracking process rather than using trained descriptor before tracking begins. The proposed updating algorithm adapts to the undergoing changes in appearance. The results demonstrate that the tracking method is capable of tracking moving targets with appearance variations in non-stationary camera video streams.