Multiple Objects Tracker Using Data Association Optimization Algorithm

With the development of the machine vision, more and more researchers in the world pay attention to multiple objects tracking technology. This technology shows great importance in many applications, such as video surveillance, human-computer interaction and intelligent system, etc. This thesis mainly focuses on the persistently tracking problem of multiple objects in clutter environments. We foumulte a network flow framework as the tracking process of the multiple objects, and then associate these objects which are in the discontinuous frames using data association method in the proposed framework. Meanwhile, the rationality of the proposed framework and the data association optimization algorithm in the design processing is analyzed qualitatively and quantitatively. The work of this paper can be divided into several aspects as follows:Firstly, construct a framework which can describe the multi-object tracking processing better. This model can satisfy the dynamic relationship among targets, help the representation of tracking process of multiple objects which is also the theoretical support of the framework under cullter environments.Secondly, realize persistent tracking of multiple objects in clutter environments by using a novel data association optimization algorithm under the proposed framework amd give the theoretical support for this algorithm. This novel data association optimization algorithm can improve the tracking performance and increase the tracking speed. It also can apply in real-time applications. Thirdly, find a method to reduce the learning time of multi-object tracker in the preprocess, and avoid the unnecessary tracking loss. There are four main innovative points in this thesis:Firstly, a novel network flow framework for objects-tracking with the min-cost is proposed. Compared with the original network framework, the new framework not only can describe the process of target moving clearly, but also can reflect the relationship of neighborhood locations between the neighborhood frames by conditional probability. The consistency between the solutions after the integer programming linearization is proved. The global optimal solution can be obtained by the data association optimization algorithm in the proposed framework.Secondly, a multi-objects tracker is designed by using an improved A*algorithm. This imporved A*algorith, named as DS-A*algorithm, not only can improve the solving ability, but also can obtain the global optimal solution. The theoretical proof is also given in the thesis. The experiment shows that, compared with the state-of-the-art data associate optimization algorithms, the performance of the proposed multi-object tracker using DS-A*algorithm is better.Thirdly, a multi-object tracker is designed by using an improved SPFA algorithm. This fast dynamic SPFA algorithm, named as FDSP algorithm, can offset the deficiency of DS-A*algorithm which consider less about the node states in the shortest path, and the gobal optimimal solution can be obtained theoretically. The experiment shows that, compared with the multi-object tracker with the DS-A*algorithm, the performance of the tracker using FDSP algorithm improves further.Fourthly, a novel preprocessing method using multi-frame overlay and batch processing is proposed. The long sequence is splitted into lots of segments, while the segments are processed in the same time, and the false negatives in the split point are considered. This method reduces the learning time of the association algorithm, and helps the real-time operation.