Research On Object Tracking Algorithm Based On Siamese Network

As an important research topic in the field of computer vision,visual object tracking technology has high commercial value.For example,it plays an important role in military and civilian fields such as unmanned driving,robots,navigation and guidance.In recent years,many achievements have been made in object tracking algorithms.However,it is still a challenge to achieve an accurate and robust tracking effect due to the complexity and changeability of real tracking scenarios.The emergence of the siamese network has achieved a good balance between the accuracy and speed of object tracking.But when the target encounters complex situations such as occlusion,rotation,flipping,and illumination changes,the tracking effect is significantly reduced.Since the first frame is fixed as the template,the feature similarity decreases during long-term tracking,which may easily lead to tracking drift.Based on the siamese network object tracking algorithm,this thesis makes improvements in feature extraction and fusion,template update,which significantly improves the tracking accuracy and robustness of the algorithm.The main work as follows:First,a multi-feature pooling fusion strategy and a multi-attention fusion method are proposed.The image features extracted by the Res Net50 network are pooled and fused at different levels.The multi-attention fusion method is used to enhance the single attention feature.We have improved the model's ability to judge the spatial position of the image.Through the improvement of attention mechanism and other modules,the accuracy and robustness of the algorithm are improved.In order not to increase the amount of model parameters after increasing the network structure,this thesis proposes an embedded asymmetric convolution kernel.It greatly reduces the amount of model parameters and ensures the real-time performance of the object tracking algorithm.Secondly,this thesis proposes the concept of template pool to improve the problem that the original siamese network only uses the first frame as the tracking template.The multi-peak detection strategy is used to determine whether the current tracking result is reliable,so as to realize the adaptive update of the template pool.It solves the problem that the matching degree of the initial frame template decreases with the passage of time,and can better deal with the situation of the tracking target's state change.Finally,in the feature extraction process,a multi-layer feature fusion module is proposed to effectively fuse deep features rich in semantic information with shallow features rich in detailed information,which improves the representation ability of convolutional features.In the case of interference from similar objects,the improved algorithm can track the object more robustly.In order to objectively evaluate the effectiveness of the algorithm in this thesis,we use the object tracking benchmark test platform to test the algorithm and compare it with the classic algorithm in the same environment.We visualized the object tracking results,and we can see the effect of the algorithm in this paper more intuitively.