Visual Target Tracking Based On Particle Filter

Visual target tracking is a key problem in computer vision, which has a wide range of applications in military guidance, visual surveillance, visual navigation of robots, human-computer interaction, and medical diagnosis, etc. The goal of visual target tracking is to enable the computer to imitate the motion sensibility of human vision, perceive the moving target in a sequence of images, and then provide an important data source for visual analysis and understanding. It often becomes very difficult for visual target tracking due to the randomicity and complexity of the target motion in real environment, such as different target sizes, the irregularity of moving target trajectory, and the similarities between target color and background color, etc. As the current researches and the algorithms can't effectively solve some difficulties in visual target tracking, it's of great significance in theory and practical applications to study target tracking under complicated background.In this dissertation, based on the existing computer vision algorithms, new algorithms are proposed for the tracking and recognition of moving targets, especially human, vehicles and hand gestures, in the applications of intelligent surveillance and human-machine interaction systems. The main contributions are summarized as follows:1. For tracking moving targets in the intelligent surveillance system, an advanced particle filter, called windowing particle filter, is proposed based on the general particle filter, to improve the tracking performance by using window-based filtering method to update particle sets and by refreshing the size of particle sets according to the target state estimation. In the novel algorithm, posterior distribution is represented by mixtures of particle sets inside estimation window. The number of particles respectively sampled from different particle sets is in proportion to the weight of each particle set, and particle states are improved by integrating the latest observation to realize object tracking. The proposal algorithm can reduce the computational cost while keeping certain tracking accuracy.2. In allusion to video moving objects in complex background, a particle filter approach that fuses two cues, i.e. color and motion, is presented, in which a layered sampling strategy is embedded. This method overcomes the instability caused by using a single measurement source, and effectively resolves the tracking difficulties due to imprecise silhouette and/or color models. The proposed fusion method can robustly track moving objects of interest in complex background.3. For the dynamic hand gesture tracking and recognition in human-computer interaction system, a novel trajectory tracking and recognition algorithm is proposed, which combines a bi-directional deep neural network called "Continuous Autoencoder" into a particle filter. First, the "Continuous Autoencoder" network embeds the high-dimensional trajectories in a two-dimensional plane based on a peculiar training rule and learns a trajectory generative model by the inverse mapping. Then a series of plausible trajectories are generated by the trajectory generative model. In the tracking process, the generated samples from the plausible trajectory set are weighted by the color likelihood and are resampled so as to obtain target state estimation at each time step in spirit of particle filtering. The trajectory identity is inferred by evaluating the improved Hausdorff distance between the estimated trajectory up to now and the truncated reference trajectories. In particular, the advantage of the method is to realize tracking-while-recognition, i.e., trajectory recognition offers dynamic probabilistic priority to guide tracking for the next time, while tracking supplies its results to trajectory recognition.