Research And Implementation Of Single Object Tracking Algorithm Based On Siamese Network

Visual object tracking is one of the most sought-after research topics in computer vision,which has a broad extent of applications,including intelligent video surveillance,augmented reality and autonomous driving.In recent years,with the development of deep learning,many tracking algorithms based on Siamese Network have achieved promising performance.However,due to a series of challenges caused by object deformation,illumination change,etc.,these methods are hard to strike a balance between accuracy and efficiency,Therefore,how to design and establish an efficient and accurate tracker is still a very challenging problem.In this paper,from the perspective of accuracy and efficiency,on the basis of fully investigating the current status of research at home and abroad,a number of issues have been investigated for visual object tracking.The main work and innovations of the paper can be summarized in three-fold:(1)A similarity re-weighting module is proposed.In the case where the target is difficult to be distinguished from the background,the similarity matrix calculated by the previous Siamese-based tracking algorithms offers few discriminative cues,which makes it difficult to accurately locate the target position from the cluttered background.In this paper,we leverage the correlation of the similarities to re-weight the original similarity matrix,which enhances the discriminative power of the similarity matrix and effectively distinguishes the target object from the background.(2)A distractor-aware tracking strategy is proposed.Previous tracking strategies only used the position of the main peak in the response map to locate the target,but the main peak may correspond to the distractor,which tends to result in target drift.In this paper,we generate sparse candidates at multiple peaks and combine independent scores of two classifiers for comprehensive scoring and ranking,effectively eliminating distractor with similar appearance.(3)A lightweight and high-performance tracking system is implemented for practical application.The system has four advantages,including small model size,high running speed,powerful generalization ability and high accuracy,which can greatly meet the practical application requirements.