The Visual Tracking Algorithm Based On Random Forest Research And Application

Visual object tracking has drawn increasing interests in recent years, mainly due to its wide variety of applications, e.g., video surveillance, human computer interaction, robot navigation, motion-based recognition and much beyond. Hence, visual tracking is an appealing research topic closely combining theories and practices.The thesis specifically focuses on adaptive appearance model for visual tracking. To begin with, the state-of-the-art trackers as well as the classical ones are reviewed. Afterwards, we propose two visual tracking algorithms, employing random forest to construct discriminative appearance models. By such a paradigm, tracking is treated as a binary classification problem. With online updating for the classifiers, proposed trackers are able to cope with appearance variations and background clutter.Firstly, a novel tracker under ensemble tracking framework using random forest is proposed. Pixel-wise features are extracted from the target area and its nearby background to initialize an ensemble of randomized trees. The classifier then labels pixels in adjacent frame either belonging to the target or the background, generating a confident map by which a mean-shift iteration is employed to locate precise target position. Meanwhile, the appearance model is updated through growing new trees to substitute those degraded ones. Benefiting from the noise insensitivity and operation efficiency of random forest, the tracker runs in real-time and remarkably outperforms the AdaBoost-based ensemble tracker in performance.Then, we propose an incremental random forest for scenarios where learning samples arrive sequentially. The proposed algorithm combines on-line bagging and incremental tree growth. Experiments demonstrate that the classification accuracy is close to its off-line counterpart and superior to on-line boosting on both synthetic and real data sets. Accordingly, the incremental random forest is applied to visual tracking as core learning algorithm under particle filter framework. The classifier is updated by sequentially arrived samples during tracking process. In addition, the target is described by multi-features, which are weighted based on their spatial uncertainty via particles. Convincing results demonstrate that the tracker manages to handle unforeseeable appearance variations and background clutters, comparable with several state-of-the-art trackers.Finally, an active tracking system is developed using a PTZ camera for face tracking. The object of interest is simultaneously detected and tracked via an ensemble tracker combined with off-line trained face and skin detectors. Experiments exhibit the effectiveness and robustness of the system.