Research On Single Visual Object Tracking Algorithm

Visual object tracking is a trending research topic in computer vision,it aims to achieve accurate tracking of a target throughout the consecutive frames of a video sequence,based on the ground truth annotations in the first frame.Despite notable progress that has been made so far,visual object tracking remains a demanding task due to the complexity and variability of the target motion patterns and the tracking environment.Building upon prior research,this thesis focuses on addressing the difficulties and challenges encountered in single visual object tracking by conducting research and making improvements to two mainstream object tracking algorithms: correlation filter and siamese network.The primary research work is summarized as follows:1.To address the problem of degraded tracking performance in correlation-filter object tracking algorithms under background clutters and occlusion challenges,an object tracking algorithm based on multi-feature fusion and occlusion detection mechanisms is proposed.The algorithm introduces background context information during the construction of the correlation-filter tracker,enhancing the tracker’s discrimination ability between the tracking target and the background.Different feature fusion strategies are designed to fuse the hand-crafted features with deep features to enhance the feature’s representation capability.An occlusion detection mechanism is introduced to solve the problem of tracking failure under occlusion challenge.The experimental results from the performance evaluation on the OTB-100 dataset demonstrate that the improved algorithm obtains a 32.80% improvement in tracking success rate compared to the baseline algorithm KCF,a 12.30% improvement compared to the CF2 algorithm,which utilizes deep features as well,and a 4.80% improvement compared to the Staple＿CA algorithm,which also introduces background context.2.Aiming at the problem of insufficient utilization of feature extraction network,inadequate feature representation capability,and inaccurate tracking in complex challenges caused by using a fixed template in siamese network object tracking algorithms,an object tracking algorithm based on feature enhanced and dual template update strategy is proposed.The algorithm first modifies the feature extraction network,and fuses features of different layers,then the fused features are passed through the spatial-channel attention module to boost their representation capability.A dual template update strategy is designed to achieve flexibly and efficiently template updating during tracking,which tackles the complexity and variability of the environment in the tracking stage effectively.Compared to the baseline algorithm Siam FC,the tracking success rate and the accuracy of the improved algorithm on OTB-100 obtains an improvement of 15.16% and 12.20%,respectively.The expected average overlap rate,robustness rate,accuracy rate and overlap accuracy rate on VOT-2017 obtain an improvement of 23.03%,17.88%,9.95% and28.06%,respectively.3.Aiming at the problem of decreasing tracking performance under low resolution challenges,an object tracking algorithm that introduces feature channel interactions is proposed,and the TensorRT-based inference acceleration for the improved network is implemented successfully.The proposed algorithm analyzes the effect of padding on the network and its offset impact on the prediction of target center position,and a cropping module is introduced to the network to eliminate the offset effect caused by the padding.A feature channel interaction module is introduced to achieve interactive fusion of features information across different channel,enhancing the feature representation capability.Then,Focal Loss is utilized as the loss function to address the problem of imbalanced numbers of training positive and negative samples,as well as the unbalance contributions between simple negative samples and hard negative samples.The algorithm achieves an improvement of 10.39% in the tracking success rate and 11.27% in the tracking accuracy on the OTB-100 dataset,in particular,it achieves significant improvement in the low resolution challenge and motion blur challenge.Finally,this algorithm conducts a research on TensorRT-based network acceleration method,and the TensorRT-based inference acceleration of the proposed improved network is implemented successfully.