Visual Tracking Based On Particle Filter And Kalman Filter Under Complex Environments

Visual target tracking is a key problem in computer vision, it has a wide range of applications in military guidance, visual surveillance, visual navigation of robots, medical diagnose, and etc. The goal of visual target tracking is to imitate the motion sensibility of human vision, empower the machine with the ability of perceiving the target motion in the image sequence, and provide an important data source for visual analysis and understanding. Although visual target tracking has been widely researched and many effective algorithms have been proposed, the robust is still the focus of the researches of visual tracking.In this dissertation, the research of visual racking in complex environment is focused on designing reasonable and efficient observation model and motion model, occlusion, and resampling algorithm, in which some methods are proposed. The main contents and contribution of this dissertation are as follows:1. The existed visual racking algorithm is analyzed, the Bayesian theory is studied including Kalman filtering and particle filtering, then the meaning of each module of the tracking framework is explained in detail, both advantage and disadvantage is discussed.2. To realize robust tracking in complex environment, a new visual tracking algorithm based on color feature and Bayesian estimation is proposed. With the current system observations obtained by particle filter, Kalman filter was used to perform update for the mean and associated covariance of the particle set, and design an optimal proposal Gaussian distribution, then new particle set from the distribution is sampled and particle filtering is performed, an occlusion algorithm with two categories was presented, by which target was divided into blocks and detected respectively and then classified into partial occlusion or complete occlusion.3. To obtain more accurate and efficient observation model, a novel visual tracking algorithm based on color feature and histograms of oriented gradients is proposed, firstly, histograms of oriented gradients are extracted to represent the moving object, then with the color feature and Bayesian algorithm framework, the algorithm performance is robust.4. To eliminate the impoverishment problem of standard particle filter, Genetic Algorithm with Markov Chain Monte Carlo mutation operator based resample algorithm is proposed, which can effectively steer the set of particles towards regions with high likelihood to prevent the impoverishment problem, and keep the diversity of particle set.