Research On Key Techniques Of Particle Filtering-Based Visual Object Tracking

Tracking moving targets in video sequences is an active research subject in the field of Computer Vision until now. Object tracking contains many technologies from different subjects such as, Image Processing, Pattern Recognition, Statistics, Machine Learning, and Biology. So it is an interdisciplinary technique. As a fundamental application, object track-ing can be found in many areas from daily life to military. Especially, it is used extensively in monitoring systems, traffic system, Human-Computer Interaction, and automotive naviga-tion system. Precisely tracking objects is the precondition of recognition and decision, and it is the basis of advanced visual system. This dissertation discusses some key technologies about the method of particle filters-based object tracking.Firstly, we propose a novel object appearance model with spatial information by using camera model and perspective projection principle, which is called Spatial Information-based Object Model (SOM). This model contains the relationship between a realistic three-dimensional scene and the image plane captured by camera. Compared SOM with the tra-ditional object appearance model, the proposed model can automatically adjust the size and orientation of search window according to the location of the object. The model is suitable for video surveillance applications in which view angle is fixed and camera parameters are known. To verify the applicability of the SOM model, the proposed model is applied to particle filtering-based pedestrian tracking, and it is also used for rotation estimation with SIFT feature and Least Square Estimation method. The experiment verifies that SOM-based particle filtering has higher stability and accuracy than traditional particle filtering, and the SOM model achieves rational results on pedestrian rotation estimation. The basic ideas and methods on estimation of (pedestrian’s) height and pedestrian detection with SOM model are given at the end of Chapter3.Secondly, to handle partial occlusions in object tracking, we propose a global and local dynamic system-based method for object tracking, which is called LDMPF (Local Dynamic Model-based Particle Filtering). In LDMPF method, GDM (Global Dynamic Model) and LDM (Local Dynamic Model) are modeled on integral and local feature from appearance of the object (rectangle region). For predicting the state of object by the global and local particles sampled from GDM and LDM, GPF (Global Particle Filtering) and LPF (Local Particle Filtering) are adopted in LDMPF. State evolution of global and local particles from GDM and LDM simultaneously can improve the accuracy of object tracking. In our experi-ments, LDMPF method is compared with traditional particle filtering and popular methods including (?)1Tracker, IVT, and FragTrack. Experimental results verify that the stability and accuracy of LDMPF method, and LDMPF can handle the partial occlusion and drift effi-ciently.Finally, to overcome the disadvantages of the traditional template matching for object tracking, we propose a dynamic template with local feature, which is called LDT (Local Dynamic Template). LDT template divides the overall template into local templates (blocks). LDT template has new strategies for template matching, template prediction, template updating, and occlusion detection. During the object tracking, each local template (block) will be updated independently. In order to prove the validity of the proposed LDT, we apply it into particle filtering framework for object tracking (named LDTPF, LDT-based Particle Filtering). The reference template in LDTPF method is generated by template prediction from LDT. According to partial occlusion and appearance changing, LDT can real-time update partial block. In experiment, LDTPF method is compared with traditional particle filtering and several popular methods. The experiment results indicate that the proposed template method has higher stability and accuracy, and can handle partial occlusion and appearance changing in object tracking application. We also combine LDT with SOM and LDM, respectively. The experimental data from SOM (with LDT) and LDM (with LDT) show that LDT method has strong applicability.