Online Learning Tracking Based On Bidirectional Corner Optical Flow And Fusional Template

Object tracking is a fundamental topic in computer vision, and is widely applied to intelligent surveillance, video retrieval, human machine interface, robotics navigation and digital entertainment. However, it remains a great challenge to achieve robust consecutive trackingunder cases of object occlusion, disappearance and reappearance, geometric distortions, similar object interference, fast movement, illumination variations, background clutters, image blur, and real time requirement.This thesis introduced the state-of-the-art techniques in visual tracking domain. It put emphasis on the methods of optical flow based object tracking, feature-based object representation, and randome forest based object detection. The major contributions of this work are highlighted as follows:1. A novel optical flow based tracking algorithm is proposed based on the traditional Lucas Kanade optical flowmethod where the tracking accuracy is improved by selecting high-confidence keypoints and optimizing tracking failure detection mechanism. As a result, this algorithm can deal with the complex scenes of noisy points, occlusions, and fast movements.2. A feature fusion template is proposed to provide a good representation of objects based on statistical features and local Haar-like features. Then a online learning based tracking model is designed using the feature fusion template for robust tracking guidance.3. A long time surveillance system is designed for consecutive object tracking in complex circumstances, where trajectory smoothness constraint is exploited in a track-detect-learn structure as well as the feature fusion template. It can achieve promising prediction of object location between successive frames.The above-mentioned works are evaluated on the public video datasets and real video sequences captured by the camera network of our institute. The experimental results demonstrated that, bidirectional corner optical flow based tracking is better than the traditional Lucas-Kanade optical flow tracking method while handling noise, occlusion, and fast movements. Moreover, online learning based tracking model with feature fusion template for object representation can achieve stable tracking under the cases of disappearance/reappearance, heavy occlusion and light changing conditions. Finnally, trajectory smoothness constraint can further improve the tracking accuracy for relay tracking task and multi-object tracking task, as well as meeting the real time requirement of a surveillance system.