Research On Object Tracking Method Based On Sparse Representation

In recent years, target tracking theory has been studied and developed widely,and it has been successfully applied in many fields both home and abroad. While there are still some questions exist in the current target tracking algorithms. For example, the accuracy and robustness of algorithms are not high when the target in the pose changes, illumination changes, fast motion, even partial occlusions.Therefore, it is still difficult to improve the accuracy and robustness of the algorithms.In this paper, we proposed a target tracking algorithm based on the combination of HOG features and sparse representation. In the framework of particle filter, the proposed algorithm applies HOG features to describe target and integrate the edge information and image information of target features firstly, and then update the target template and classifier dynamically after using sparse representation to construct the appearance model of the target. The main components in the proposed algorithm are prediction and tracking modular. Firstly use particle filter integrated HOG features to predict the target position of next frame, and then calculate the confidence value of each predicted position by Naive Bayes classifier to track target accurately. Although the proposed algorithm has a slightly high computational complexity, it improves the robustness and accuracy in the tracking system to some extent.In order to verify the tracking performance of the proposed algorithm, we set 5groups of experiments to compare our algorithm with IVT, L1, PCA and MIL classic algorithms. The first group of experiment aims at comparing the four algorithms under the interference of partial occlusions and pose changes, while the main factor considered in the second group is illumination changes. The third is used to observe tracking performance of the four algorithms in the complex background and movement mutation. The fourth is carried out to test the robustness of four algorithms when the target is occluded by the similar object. The last is designed to detect the disturbance of fast motion and blur target for the four algorithms. The experiment results show that the proposed algorithm performs better in pose changes,illumination changes, occlusion, fast motion and blur target than IVT, L1, PCA, MIL algorithms, but its performance is not as good as MIL when target moves suddenly.