Visual Tracking Based On Particle Filter Under Complex Observation Conditions

With the development of computer science, electronic technology, automatic control and artificial intelligence, the intelligent visual surveillance promises a wide range of applications in many social areas, such as national defense, human-computer interaction, security and space exploration. So people pay more and more attention to the researches of visual tracking, which is a key technology of the intelligent visual surveillance and is a very important research topic in computer vision. Bayesian theory provides a good framework for integrating prior knowledge and maximum likelihood estimation. Bayesian based particle filtering technology has been proven to be successful for non-linear and non-Gaussian estimation problems and effective for solving tracking problems. Within methods of this class, tracking is modeled by a state-space time series estimation problem, which is solved using a sequential Monte Carlo estimation method. The major problems in designing particle filter based visual tracking algorithms include: eliminating particle impoverishment, designing reasonable and efficient observation model and motion model. Generally speaking, obtaining an accurate motion model is difficult by using only images. So people are more interested in the researches of observation model and the impoverishment problem. In fact, many factors can affect the effective observation of the tracked targets in images, such as the complex background, multiple targets, occlusion, pose change, fast motion and so on. People hope that the targets can be tracked stably and persistently. So, the robust performance is the focus of the researches of visual tracking. We aim at resolving the current problem of visual tracking in this paper, in which some methods are proposed to realize the robust tracking using particle filter under complex observation conditions. The main contents and contributions of this dissertation are as follows:1). To realize multi-target tracking in outdoor and complex environment, an adaptive targets extraction based tracking algorithm using the augmented particle filter for image sequences is proposed, which is Gaussian mixture model based. Firstly, a Gaussian mixture model is built to model the background of the monitored environment. Then we utilize pixel changing detection, shadow detection and morphological operator to detect targets. In the tracking phase, we use the augmented particle filer to track targets. It uses the Kalman filter to take advantage of the most recent observations, and the computing of observation likelihood is used to process target entrance, target exit and mutual occlusion.2). To resolve the problem of target tracking in the circumstance of illumination and pose changing such as rotation, shift and scaling. We propose an adaptive observation model based particle filter tracking scheme. Firstly, the tracked target is modeled by an adaptive Gaussian sequential density approximation, which can adapt the change of illumination partially. Secondly, robust statistics is introduced to handle the occlusion in computing observation likelihood, and affine transformation based similarity measurement is used to tackle the change of pose of target.3). To track the fast target, we propose an improved particle filter based visual tracking method. Firstly, we use Gaussian kernel function to replace the Dirac delta kernel function, and embed the re-sampling into filtering to eliminate the impoverishment problem of the standard particle filter. To improve the performance of particle filtering further, we introduce the genetic evolution operator into the process of Gaussian kernel particle filtering, which can effectively steer the set of particles towards regions with high likelihood and prevent the impoverishment problem. The proposed method can track the fast target effectively using fewer particles than the standard particle filter.4). Visual contour tracking in complex background and the pose changing of target is a difficult task. The observation model is often nonlinear in images. Traditional visual tracker based on particle filter requires a large number of particles and often collapses during the tracking process. This paper presents a new contour tracker based on the improved particle filter that is superior to the standard particle filter in many practical situations. Firstly, the tracked objects are modeled as B-spline curves, and the planar affine transformation is used to tackle the pose changing of the targets in the shape space. Secondly, it uses the maximum likelihood estimation method to learn the motion model parameters and employs a more accurate nonlinear observation model. Finally, it uses the re-sampling based on genetic evolution. Performance comparisons show that the proposed method is an improvement over the traditional particle filter.