Research On Multi-target Track Association Algorithm Based On Random Finite Sets

As the key technology in data fusion system, multi-target tracking has received a lot of attention and researched by domestic and overseas scholars, which played a very important role in modern military and civilian fields. The application of the random finite sets theory in the fields of multi-target tracking has been in rapid development in recent years. It avoids the data association technologies in traditional target tracking and brings new development in multi-target tracking. However it does not provide tracks of individual targets, makes it difficult to deal with the subsequent data fusion. How to associate the tracks in multi-target tracking based on random finite sets is a topic worth studying. This thesis mainly focuses on the various methods of multi-target track association based on random finite sets. Main contributions are as follows:Firstly, several classical multi-target tracking algorithms based on random finite sets such as probability hypothesis density filter and cardinalized probability hypothesis density filter and their implementations of Gaussian mixture and particle have been introduced. The advantages and disadvantages of two kinds of filter are discussed and the evaluation standard of multi-target tracking based on random finite set is analized in this paper. This section mainly makes a good foundation for the following research in the rest of paper.Secondly, for the problem of state extraction is not accurate for GM-PHD filter when multiple targets closely spaced,a collaborative penalized GM-PHD filter has been proposed, and renormalization schemes to refine the weights has been employed to improve the estimate accuracy. Then combined with labeling PHD algorithm, the normalized correct associations has been improved for close targets. Because the target number estimation has a direct influence on track association, while the estimated target number of PHD filter is not accurate, for this issue, this paper combines the labeling CP-PHD with CPHD, the track maintenance performance has been improved.Lastly, for the problem of state extraction is not accurate for particle-PHD filter,a new state extraction method has been improved, combined with estimate-to-track algorithm,the performances of both the target states and the normalized correct associations have been improved. An improved track association algorithm is proposed in this paper to overcome the matter of wrong association when multiple dense targets closely spaced. A distance matrix is proposed and constructed to identify the targets location information to each other, and the predictions of target trajectory are associated with the corresponding tracks directly of dense targets in close proximity. It provides correct association for close targets. This proposed algorithm has great robustness and strong anti-disturbance ability. The simulation experiments show the effectiveness of the algorithm.