Research On Object Tracking Algorithm Based On Multi-algorithm Fusion

Object tracking through image sequences is the main component in the field ofcomputer vision. It has been widely applied in the national field and military affairs,such as medical diagnosis, intelligent transportation system, ballistic missile defense,military guidance. At present, particle filter is the main tracking algorithm, but it hasthe particle degeneration defect. Therefore, the important research topic in the objecttracking through image sequences is how to resolve the particle degeneracy andparticle impoverishment problem. In this paper, the target state estimation theorywith nonlinear and non-gaussian system as the main line, based on traditional targettracking method, research combined particle filter theory and mean shift theory arestudied systematically. Aiming at tracking algorithm optimization and the problemsof multi-object tracking such as objects collision, merging and separation, we studyand put forward a novel algorithm to improve tracking speed and robustness in thecomplex condition.The following is done in this thesis:(1)An adaptive immune optimization Unscented Particle Filter (AIO-UPF)algorithm is proposed. The algorithm uses the global optimization ability anddiversity of features of the immune algorithm in the affinity and concentration andthe Metropolis criteria with the adaptive threshold factor δ makes the particle setmove towards the higher likelihood area. In this way, the diversity and effectivenessof particle has improved and the problem of particle degradation and depletion hasalleviated. Simulation results indicate that this Particle Filter (UPF) algorithm hashigh performance of the state estimation.(2) A multiple targets tracking algorithm based on AIO-UPF and MS is given.The new algorithm confirms the target state sequence used the associated matrixwhich builds with the multiple targets tracking algorithm based on reasoning afterextracting target observation value. Then, using the character of anti-overlap ofAIO-UPF and real-time of Mean Shift, the frame number factor τ and handingmechanism of target conflict are introduced, tracking system can adaptively selectthe unscented particle filter algorithm of adaptive immune optimization and meanshift algorithm for multi-object tracking state estimation. Simulations show that thealgorithm can effectively solve the problems of target block conflict, merger and separation so as to achieve multi-objective robust tracking in actual scene.