Research On Algorithm Of Object Tracking Based On Generalized Hough Transform

Object tracking in video sequence is a popular topic in computer vision area. The use of object tracking is pertinent in the tasks of:automated surveillance, video retrieval, auto-navigated vehicle, human-computer interaction, sport live broadcast, etc. The general researching path of object tracking is to simulate human vision behavior and to imitate the inference mechanism of human being. Great quantities of approaches, from visual feature to analysis method, are proposed on this researching path. As we can see, numerous algorithms have been presented on some particular applications, and have been proved to be effective. However, common solutions or uniform frameworks for most problems are still not available, due to the complicated circumstance of real world and the changeability of tracked object itself. Some crucial challenges need to be settled, such as background noises, illumination changes, inter-object occlusions, object’s appearance and pose changes.After thoroughly studied on the art-of-state object tracking algorithms, this dissertation intensively investigated applications of Generalized Hough Transform in object tracking field, discussed some key issues on this topic and proposed a series of effective solutions, including classifier design, visual feature selection, online model updating, location estimation, occlusion reasoning, and multi-object association, etc. The major achievements include:(1) Semi-supervised Hough Forest Tracking Method. Proposed a flexible semi-supervised learning method based on Hough forests classifier. Introduced a random label distribution, which is based on spatial consistency in context and object-specific information detected in tracking procedure, to improve the performance of classifier, and reduce confusion between inner-class objects. Meanwhile, a Particle-Filtering-Kind random sampling scheme was implemented in detection and updating phases, which can accelerate the whole tracking procedure.(2) Detection feedback based long-term tracking algorithm. Introduced close-loop idea into tracking field. Used optic-flow based tracker as the feedforward controller, and a detector based on complementary features as feedback controller. Then the latter one’s result was used to adjust the tracker’s parameters and to maintain the appearance model of the tracked object. This feedback scheme dramatically increased the robustness and stability of tracking system on long-time axis.(3) Multiple objects tracking with Dual-level Particle Filter Embedded Semi-supervised Hough Forests. Used a dual-level particle filter, which disposed patch random sampling and position random sampling processes in one uniform sampling framework, raised time performance of the detector. And an objects maintenance scheme was designed out to cope with occlusion, background changing, entry/exit and so on.(4) Multi-pedestrian real-time tracking with multi-view multi-part Hough Forests Detector and on-line Conditional Random Field(CRF) model. Divided pedestrian object into multiple parts, learned full body appearance models and parts appearance models respectively based on online Hough Forests algorithm, which improve the model’s robustness and stability against occlusion. And a CRF model was employed to formulate dependencies and affinities between tracklets and detection responses in a short time sliding window to represent current global association, which could handle occlusion and confusion in complex scences well.