The Research Of Intelligent Multi-target Tracking Algorithim Based On Particle Filter

The multi-target tracking technology is a hot issue in computer vision, which has broad application prospect in many fields of modern society. Compared with single target, the situation that the multi-target tracking technology need to deal with are more complex, including the real-time detection of multi-target, interactions and occlusion the variable number of the real-time target, etc. Particle filter algorithm is a kind of nonlinear Bayesian filter with multi-modal search capability, which provides a second optimal solution when apply in nonlinear and non-Gaussian space. It offers a new perspective for the research of multi-target tracking.Therefore, the relevant research work in the thesis is focus on the proposed multi-target tracking algorithm that based on the particle filter. For the shortcoming of these algorithms, corresponding improvements are proposed and a more effective intelligent multi-target tracking algorithm is presented.The main works in the thesis can be divided into three parts:Firstly, based on the study work of the theory of particle filter and its applications in multi-target tracking, some improvements are proposed to overcome the shortcoming existing in these multi-target tracking algorithms. Furthermore, an intelligent multi-target tracking framework is constructed based on these methods.Secondly, for the optimal design problems in multi-target detection, an algorithm combined the HOG feature with Gentle Adaboost detection is proposed to improve the effect of multi-target detection. The result of the relevant experiments shows the superiority of HOG feature in pedestrian detection and the Gentle Adaboost is more stable than the other kind of Adaboost algorithms respectively, and it is more effective with the combination of both methods.Finally, target motion model and likelihood model are optimized designed in tracking part. For the design of likehood model, a multi-feature fusion strategy of combining the sub-integral histogram with the LBP feature is used to construct the likehood model of target in this thesis. With these improvements stated above, experimental results shows that the intelligent multi-target tracking algorithm based on particle filter proposed in the thesis is not only able to detect targets intelligently but also better adapt to challenging situations of occlusion and the cross movement of targets in multi-target tracking environment.