Visual Object Tracking Method Based On Deep Learning

With the rapid development of high quality image acquisition equipment and deep learning,computer vision has gained significant progress in the field of visual tracking.Visual tracking is an important part of computer vision,which is widely used in the fields of unmanned vehicles,video surveillance,and human-computer interaction.In recent years,deep learning have become the mainstream methods in this field,and the results validated by several large publicly available datasets show that these methods can combine accuracy and robustness.In this paper,we mainly analyze and study the single target tracking method based on Siamese networks,optimizing and improving on the current neural networks,and the main research points and innovations are shown below.(1)The Siamese network focuses on the semantic information in the video sequence:when the semantic information in the video sequence is more significant,the tracking effect is also better;and when the semantic information in the video sequence is insufficient,the tracking effect is also greatly affected.Therefore,in this paper,we consider combining low-level features such as color features and directional gradient histogram features to increase the weight of the target's surface color information,texture information and edge information in visual tracking to supplement the low-level details of the target and obtain more accurate and robust results.(2)By observing the Siamese network has different sensitivity to the target in different network channels during tracking,and use an attention mechanism to simplify the redundant features in the network and use the remaining highly relevant features for target tracking.On the other hand,this paper uses APCE to make confidence judgment on the target,and only after the target passes the confidence judgment,the information of the target will be weighted into the template,otherwise the frame is directly ignored.In this way,the template update mechanism can be introduced to improve the accuracy and robustness while preventing the template drift.This paper applies the above research work to the Siamese network,and compares the performance of the original methods on major public datasets to prove that these improved methods are effective,so that the original Siamese has been significantly improved in processing fast motion and illumination changes,etc.