Research On Single Object Visual Tracking Algorithms Based On Siamese Network

Single object tracking has an extensive range of applications in intelligent monitoring,human-computer interaction and driverless car.Although single object tracking research has made good progress,complex backgrounds,interfering targets,and occlusions will all have a great impact on tracking performance.Therefore,single object tracking algorithms still face some challenges,including accuracy,robustness,and realtime.Because the single object tracking algorithm based on the Siamese network framework can effectively balance the accuracy and time.Therefore,this thesis mainly studies the single object visual tracking algorithm based on the Siamese network.The main research results are as follows:(1)Using the attention mechanism to improve the discriminative power of image features,a multi-source attention feature fusion network combining position,channel and cross-channel is proposed to enhance object features and search features.Among them,cross-channel attention can enhance the degree of matching between object features and search features,and multi-source attention features are organically combined through a weighted fusion layer.We further adopt an online update strategy to improve the robustness of the model.The experimental results show that Siam AFC achieves state-ofthe-art performance on several benchmarks.In particular,our model achieves the best tracking result on the GOT-10 k dataset,with an average overlap of 52.3% and a success rate of 63.2%.It also performs the best robustness result with 0.131 on the VOT2018 dataset.(2)A single object tracking model Siam AFL based on cross-correlation and anchorfree is proposed,which improves the robustness of the algorithm.First,an object feature selection strategy based on the Pr ROI pooling layer is adopted to realize the adaptive selection of object features.Then,based on the pixel cross-correlation,global correlation and attention mechanism,a feature embedding module PGA-XCorr is designed to integrate the object template features and the search region feature well.Finally,an anchor-free network using location quality evaluation and background prediction is proposed to improve the anti-interference ability of the tracking model.The experimental results show that Siam AFC achieves state-of-the-art performance on several benchmarks.In particular,our model achieves the best tracking result on the GOT-10 k dataset,with average an overlap of 57.9% and a success rate of 67.7%.It also performs the best robustness result with 0.110 on the VOT2018 dataset.Compared with the Siam AFC algorithm,the Siam AFL algorithm improves the robustness of single object tracking under the premise of ensuring the tracking accuracy.