Object Tracking Algorithm Based On Particle Filter

Video based object tracking is a very challenging problem in computer vision, owning wide applications in many areas, e.g., video surveillance, intelligent traffic control, intelligent human-computer interaction, video compression, military, etc. Object tracking is a comprehensive technique of interdisciplinary and owns significant value in both theory and practice, merging together the theories from image processing, machine learning, pattern recognition, etc. Through people’s long-term study, object tracking has made great progress. However, up to present it is still difficult to achieve accurate tracking in complex scenes. As a consequence, in this thesis we shall explore some key techniques of particle filter based object tracking algorithms. Our major contributions include:(1) We propose a j oint model of a local descriptor based generative model and a global template based deterministic model, using L2 norm minimization to perform object tracking. To handle the occlusions and object variations possibly existing in the tracking procedure, we establish proper strategies for updating the coefficient vector of the generative model and the positive-negative templates of the discriminant model, assigning strong discriminative ability and adaptability to the templates. We compare our proposed algorithm with the state-of-the-art algorithms. Extensive experimental results show that our algorithm is more accurate and stable than the competing algorithms.(2) We propose a local principal component analysis based algorithm for object tracking. Appearance model is the core procedure of tracking algorithms, as the low level information extracted from images are important for determining the possible regions where the target objects may occur. Hence, the key for advancing object tracking is to establish the appearance model that has good adaptability. The algorithm proposed in this thesis learns the low-dimensional subspace in an incremental fashion, thus adapt to the variations of the object appearance. With the framework of particle filter, our algorithm partition the target region into multiple blocks, applying incremental principal component analysis in each block, and use the combination of the similarities from multiple parts as the particle weights. Incremental principal component analysis based model has two important aspects:updating the mean of a fresh sample and using forgetting factor to increase the weights of the recently acquired images. Both aspects can improve the overall performance of tracking.