Study On Multi-target Tracking Method Based On Distributed Co-evolutionary Agents

As one of the most active research in the field of computer vision, image processing and pattern recognition, moving object tracking technology has been playing an extremely important role in military affairs, security surveillance, human-computer interaction and many other areas. At present, particle filter (PF) is one of the most commonly used methods in moving object tracking. As an excellent target tracking algorithm, it is easy, flexible, and can be used in non-linear and non-Gaussian systems. However, particle filter still has some problems. One is that the robustness of the observation model, which may be influenced by the light changes and the similar color objects in the scene when modeling the target using the single color feature. Another is the problem of particle degradation. Although the re-sampling process can solve the problem of particle degradation to a certain extent, particle deprivation will become another problem and the tracking accuracy will be decreased.Multi-agent technology is originated from the field of distributed artificial intelligence, the sociality of human intelligent activities and the collective wisdom generated by the social interaction involved stretch the limitations of the traditional artificial intelligence, and provides a new way for problems solving in computer application.For the problem existing in particle filter, the main contributions of this thesis are as follows:1. A multi-feature fusion method is proposed for target modeling. The LBP feature of the target is extracted first and then fused with the color feature for modeling the target, which can increase the information of target observation and improve the robustness of the observation model.2. The tracking agent model is built based on PF. Each particle in PF is represented as an agent having the ability of local perception, competitive selection and self-learning by the re-definition of its local living environment and its co-evolutionary behaviors.3. A novel particle filter for target tracking based on distributed co-evolutionary agents is proposed. The multi-agent co-evolutionary mechanism is introduced into the particle re-sampling process, which is accomplished by the co-evolutionary behaviors among particles such as competition, recombination, migration and self-learning, etc. It can not only ensure the particle validity but also increase the particle diversity. Experimental results show that the proposed algorithm is much more accurate and robust.