Research On Multi-object Tracking Algorithm Based On Minimum Spanning Tree Model

Nowadays, target tracking is an important research direction in computer vision field, which involves image processing, pattern recognition, artificial intelligence, automatic control and other frontier theory. It is widely used in traffic control, security management, human-computer interaction, intelligent vehicles, medical testing and other fields, which has high practical value for human life. But in the complex tracking scenes, the performance of tracking system is affected by a variety of factors, such as target occlusion, illumination change, attitude diversity and so on. Therefore, designing the tracker with good performance is a very challenging work, which has important practical significance.Aiming at the deficiencies of multi-target tracking algorithm in video sequences under complex environments, an improved tracking algorithm based on space constraints is proposed in this paper. Most trackers tend to focus on significant features of the target appearance, while ignoring the relationship among different targets. Due to the frequent occlusion and similar appearance among targets, it is prone to tracking drift. Therefore, this method incorporated space location information among the targets, combining an appearance matching score with the structure deformation score, and then maximizing the configuration score to output the best location configuration of multiple targets. In the multi-target tracking scene, the tracker can achieve more accurate tracking. The proposed algorithm treats every target as a part. Firstly, the histogram of oriented gradients(HOG) features are extracted from the target areas, then the support vector machine(SVM) classifier is combined to train samples. After that, the appearance model of each target is obtained. Based on the principle of deformable template, the minimum spanning tree(MST) model is proposed to establish the relation among the various parts. The computing complexity of the whole detection process is less and the target space structural constraints are effectively represented at the same time. During the tracking process, a structured SVM algorithm framework is used to online learning parameters. Because the appearance model of all targets and the structural constraints among these targets are updated in real time, the tracker is able to adapt to target and environment changes timely. Finally, the experimental results show that this paper can improve both the effectiveness of the multi-target tracking and the robustness and accuracy of the tracker.