Study On Adaptive Particle Filter Tracking Based On Human Vision Mechanism

As one of the core issues in the field of computer vision, object tracking aims at studying the problem of tracking moving objects in video sequences. It is the basis of video analysis and understanding. It integrates the multi-disciplinary knowledge of artificial intelligence, pattern recognition, image processing, automatic control, computer science and so on. It is widely used in intelligent video surveillance, medical diagnostics, robot vision, military and other areas. Therefore, it is of great scientific significance and boasts wide application prospects.Deterministic tracking and probabilistic tracking are two main methods for object tracking. As a typical probabilistic tracking method, particle filter is widely used in object tracking because of its good performance and strong robustness. Also, compared to traditional methods, particle filter does not necessarily require the random quantity to be constrained by Gaussian distribution when used to solve nonlinear problems. However, the complexity of the practical application environment poses great challenges to the particle filter algorithm.As the state of a moving object changes, its features will also change, influenced by complex backgrounds, illumination changes, as well as the changes in the scale and pose of the object. Therefore, it is difficult to represent various moving states of an object in a complex background through a single feature, which will result in low tracking accuracy and poor tracking robustness. To solve such problems, starting with the complementarity among various feature information and based on particle filter, this thesis proposes an adaptive feature-fusion object tracking algorithm. It can improve the tracking accuracy and enhance the robustness of the traditional single-feature-based object tracking algorithm. However, because such an algorithm needs to calculate various features of an object, the computational complexity is increased and thus a real-time tracking cannot be assured. To deal this problem, this study simulates the characteristics of learning and memory of human vision in object tracking process, proposes the object tracking method of human vision mechanism and particle filter. Specific contents of the research include:1. Describe the problem of object tracking that is based on Bayesian Estimation Theory; introduce Monte Carlo particle filter method; introduce in detail the issues of importance sampling, sequential importance sampling and particle degradation; analyze the reasons for the particle degradation as well as its impact on algorithms; provides the standard to measure the degree of particle degradation; and sum up the classical particle filter algorithm.2. Adopt the adaptive feature-fusion method using the complementarity among the advantages of different feature information based on the fact that different features are adaptive to different scenes, to solve the problem that it is difficult to accurately represent various moving states of the object in a complex background. This feature-fusion method integrates various features of the object so that the limitation of the single-feature-based method can be avoided and it can distribute weights according to the reliability of different features to improve the utilization of reliable information and reduce the impact of uncertain information. Combined with particle filter, this method can achieve a stable tracking of the object. The experiment results show that the algorithm has better performance.3. Simulate the characteristic of continuous learning of human vision in object tracking process based on human vision intelligence; build a template library of the object; simulate the characteristics of memory of human vision by way of memorizing the historical information of the object and update the memory information of the vision according to the frequency change of the object; track the object based on the particle filter. The experiment results show that the algorithm has better accuracy and robustness.