Partial Least Squares-Based Incremental PCA for Robust Human Detection and Tracking

Human tracking is a major issue in computer vision where the challenging part is to track a subject under an uncontrolled environment. Most of the existing algorithms are not able to perform well in such condition due to variation in the human appearance caused by covariates like clothing and the illumination changes. One reason these algorithms fail to perform well is because of the use of a fixed appearance model for the human object. The fixed model is limited and insufficient to cope with the constant appearance change in the image stream. Overtime, the tracking result will be drifted away from the trajectory and lost track of the actual target. In this paper, a method coined as Partial Least Squares-based Incremental PCA (PI-PCA) is proposed to address the human detection and tracking problem. A human detection method based on partial least squares (PLS) regression is used to locate the occurrence of a human subject in the video frame. Once a human object is detected, incremental PCA will be used to track this subject over the video stream. A forgetting factor is used to follow the tracking history. Once the sign of a drift is detected, PLS will be called to correct and re-lock the actual position of the target object. The proposed method is an improvement over the existing incremental learning algorithms as it introduces a corrective mechanism in the tracking process. Empirical tests demonstrate that the proposed PI-PCA method adapts well to appearance change of the human object over a long video stream with substantial motion switch.