Research On Visual Single Target Tracking Algorithm Based On Siamese Network

Target tracking has become a hot research problem in the field of computer vision,and is widely used in many fields such as unmanned aerial vehicle video tracking,intelligent camera surveillance,medical diagnosis,and robotic visual navigation.Single-target tracking theory and techniques also play an important role in military fields such as long-range strikes and enemy detection,as well as in daily life fields such as finding lost children and tracking illegal vehicles,which have attracted the attention of many domestic and foreign researchers.To address the problem that existing single-target tracking algorithms have low tracking accuracy when performing tracking tasks in the face of complex scenes such as light changes,occlusions between scenes,and camera shake,this paper proposes a Siamese network singletarget tracking algorithm based on twin feature fusion ff-Siam RPN++.In order to enable the single-target tracking algorithm to learn image features with richer semantic information,this paper proposes an improved twin feature fusion module,which performs feature fusion on the feature images processed by the twin feature extraction network of the Siam RPN++ algorithm,stitches the extracted multi-layer image features,and finally uses the pyramid feature extraction module to extract the stitched feature maps and feeds them into the subsequent RPN module for classification and tracking.This thesis compares the ff-Siam RPN++ algorithm with various single-target tracking algorithms and shows that the ff-Siam RPN++ algorithm achieves better results in single-target tracking than other single-target tracking algorithms.In order to further improve the accuracy and anti-interference capability of the Siam RPN++ single-target tracking algorithm,a hybrid attention mechanism is proposed in this paper to weight the attention of the feature map.This paper proposes an improved spatial attention mechanism after visualizing the channel-averaged activation values and analyzing the image information of the feature maps extracted by the Siam RPN++ algorithm,and then introduces a hybrid attention mechanism by combining the normalization-based channel attention mechanism with the spatial attention mechanism.The improved hybrid attention mechanism is combined with the Res Net-50 network in the Siam RPN++ algorithm to extract image features and obtain the image features filtered by the attention mechanism.Finally,the improved Res Net-50 network is combined with the twin feature fusion module to propose the single-target tracking algorithm ff-Siam RPN++-a,which takes the image features extracted by Res Net-50 and feeds them into the RPN module for target tracking after processing by the feature fusion module.In this thesis,the improved hybrid attention mechanism is ablated with the original Siam RPN++ algorithm,and the ff-Siam RPN++-a algorithm is compared with a variety of Siamese network-based single-target tracking algorithms.The experimental results show that the attention mechanism can effectively filter out the feature channels and spatial locations that are more information-rich in image features,while the single-target tracking algorithm ff-Siam RPN++-a improves tracking accuracy and robustness compared to the traditional Siamese network-based single-target tracking algorithm.