Research On Particle Filter Based Reliable Visual Tracking

Visual tracking is one of the most important applications and research topics in the field of computer vision which is extensively applied in the aspects such as Intelligence traffic system, video communication and compression, human-computer interaction,multimedia content retrieval and so on. Visual tracking algorithms are used to detect, extract, recognize and track moving targets in videos, thus to obtain the motion information such as position, speed, acceleration, trajectory .etc. This information is the basis for handling and analyzing videos, in order to understand the behavior of the interested target. The common difficulty for visual tracking is how to modify the tracking algorithms, in order to realize reliable tracking under complex visual environment which usually contains object occlusion, target deformation and scaling changing, abrupt motion .etc.As an extensively approved tracking method, Particle Filter is popular in many fields since it is suitable for non-linear and/or non-Gaussian applications. However, Particle Filter has limited performance under problems like similar-colored object occlusion, target deformation, long-time full occlusion .etc. This paper targets at improving Particle Filter to realize the reliable tracking under such complex visual environment.Similar-colored object occlusion means there are objects with similar color as the target in the background. In this condition, the performance of color-based particle filter will degrade greatly and the tracking becomes difficult. The root of this problem lies in the inefficiency of color feature in representing and distinguishing the target. Therefore, this paper improves color-based Particle Filter and designed a multiple observation models with the integration of color and gradient orientation, which make a statistical analysis on target texture besides color; also, a Gaussian weighting function based on pixel location is introduced to reduce the bad effect of outlier pixels, which increases the accuracy of target model as well as the robustness of particle match.Target deformation and scaling changing are common tracking problems, which are caused by variation of illumination and viewing angle, rotation or motion. Scaling-alterable particles and an adaptively-updating reference target model are used in the paper. As a result, particles could be adaptive to fit the tracking scene well and achieve better results.Long-time full occlusion is another difficult problem in occlusion. After analysis, it is found that particles propagate in a limited range in long-time full occlusion and thus the target might occur outside the particles'range which causes failure. In order to solve this occlusion problem, a scheme of multiple state noises based state transition model is proposed. The range of particles is determined based on a dispersion evaluation. This can ensure the redetection of the target at the end of full occlusion as well as the tracking accuracy under occlusion. Besides, Local binary pattern (LBP) feature is incorporated in the observation model, which describes the texture information of target, and improves the accuracy of tracking.Experimental results show that this paper achieved reliable tracking results based on both people sequences and car sequences, and it handles problems such as similar-colored object occlusion, long-time full occlusion as well as target deformation and scaling changing well.