Stable Object Tracking Based On Learning And Detection

The thesis studies the stable object tracking technology based on learning and detection.Its purpose is to explore the basic principles and basic mechanisms of the object tracking process,and to achieve stable and effective long-term tracking of general objects.Research on tracking methods for general objects requires that the features be more universal,and track the interest objects without prior knowledge(without pre-training).The long-term tracking of the object often occurs in various complex changes such as: lighting changes,object rotation,rapid movement,the goal of shielding or leaving the field of vision,making this issue very challenging.The thesis reviews the development of tracking methods for feature extraction and object classification framework for long-term tracking,and discusses the reasons for the failure of tracking due to changes in light illumination,object rotation,and rapid movement.From the comprehensive analysis of the literature,we can see that tracking algorithm based on TLD(Tracking-Learning-Detection)integrates object feature learning,object detection and object tracking,and shows excellent performance in long-term tracking.Therefore,this paper carries out an in-depth analysis of the algorithm,and analyzes and analyzes the reasons for its failure,including:(1)In view of the fact that the features used in the TLD algorithm are not robust enough,this paper improves the detection module of the TLD algorithm by using the good characteristics of the LBP rotation invariant mode.The experimental results show that the tracking effect is improved,and to a certain extent,the problem of tracking failure occurred when the light changes and the object rotates.(2)Aiming at the problem that the tracking failure detection mechanism in TLD algorithm seriously affects the tracking result when the object has a rapid movement,this paper proposes a tracking failure detection mechanism that adaptively adjusts the threshold according to the object and the image size.Experiments have proved that this tracking failure detection mechanism proposed in this paper can effectively solve the false tracking failure caused by the intense movement of the object.(3)For the problem that the object model in the TLD algorithm is not reasonable enough,this paper uses the confidence of the tracking results to improve the original object model.The object model after confidence weighted processing can better represent the object.Experiments show that the tracking results of the proposed method outperform the TLD algorithm in center position error and tracking success rate.This article uses public data sets to test and verify the algorithm,and compares the experiment with the TLD algorithm with excellent tracking effect as the baseline.It shows that this algorithm can effectively solve the problem of tracking failure when the object has light changes,in-plane rotation,and rapid movement.At last,the algorithm is comprehensively analyzed and summarized,and the idea of continuous improvement of the tracking algorithm is given.