Study On Object Tracking Algorithms Based On Structured Support Vector Machine

As technology advances, computer vision follows the footsteps of artificial intelligence into the visual field. The issue of object detection and tracking is a key and challenging problem in computer vision. In recent years, various scholars in the relevant areas of research focus on this field. They have found different detection and tracking algorithms responding to different environments.Between these two areas of detection and tracking, the most critical problem is identical, i.e., how to effectively describe the object, and how to make computer recognize the object accurately. At the same time, the significant differences between these two are the object detection pays attention to accuracy and the other pays attention to the real-time tracking.To deal with these two requirements, this paper adopts SVM classification which has excellent classification features. Specific researches are as follows.For object detection system, it would be divided into two phases. Phase one is the training phase. Firstly, HOG and LBPHF algorithms could be calculated respectively in each sliding window to construct the feature set. Subsequently, the linear support vector machine(SVM) would train a classifier. Finally, in order to obtain the optimal discriminant model, this paper used bootstrap method to update the classifier. The other phase is the detecting phase, during which the augmented feature set would be put into the classifier. Then the Non-maximum Suppression was adopted for overlapping detections. The experiments show that the improved method has met these requirements: the higher detection rate, lower computational complexity, and anti-interference ability of pedestrian limb deflection.For object tracking system, firstly this paper uses the model-free tracking framework. It adopts the improvement of HOG-LBPHF and the structured information between the objects to train SVM. Then the passive-aggressive perceptron is used to update hyperplane. Finally, the location of objects in the next frame would be determined by the minimum spanning tree. Through the experimental comparison, the algorithm in this paper has excellent performance.