Adaptive Visual Object Tracking Method Based On Siamese Network

Visual object tracking is one of the basic research of computer vision,which is widely used in human-computer interaction,public safety monitoring and automatic driving.The main task of visual tracking is to use the tracking algorithm to predict the position of the target in subsequent video frames with the target initial position.In an unstructured natural scene,the tracked target will inevitably encounter problems such as illumination changes,target scale changes,target fast motion,and complex background interference,which poses a huge challenge for long-term robust visual tracking.Due to the existence of target diversity and complex background interference,the research of this thesis aims to design a robust adaptive visual object tracking algorithm to solve the above problems.Based on the theory of visual tracking,this thesis studies the visual tracking algorithm based on siamese network structure,and starts the research from the two directions of multi-feature fusion and attention mechanism.The specific work and innovations are as follows:1.In order to build a visual tracking algorithm framework based on siamese network,this thesis explores the basic knowledge of target tracking.In terms of feature extraction,the implementation mechanism of the manual design features HOG,CN and CNN is analyzed.In the aspect of algorithm framework,the siamese network basic structure and correlation filter computer system are studied.In terms of visual tracking datasets,summarize the model training dataset,compare experimental validation datasets,and compare experimental evaluation metrics.2.Most feature fusion methods only perform simple linear fusion of multiple features.In order to explore the deep connection between features,this thesis proposes a feature fusion visual tracking method based on siamese network.This research is expected to establish the relationship between feature extraction and feature fusion to improve the adaptability of the tracking method.In the process of model training,the parameters in the feature extraction layer and the feature fusion layer are continuously optimized through the backward feedback mechanism,which improves the tightness and flexibility of multi-feature fusion.During the tracking process,the fusion result of the network output can more reliably predict the target location to solve the problem of target apparent diversity and complex background interference.Finally,the experimental experiments were carried out on the experimental datasets TC-128,OTB50 and UAV123.The experimental results show that the feature fusion network can effectively improve the tracking performance.3.This thesis uses high-dimensional features to model the target.In order to quantitatively distinguish the importance of the eigenvalues on different channels of the same feature to the objective superficial modeling,this thesis introduces the attention mechanism based on the feature fusion network to improve the adaptability of feature fusion on different feature channels.Finally,comparative experiments were carried out on the experimental datasets TC-128,OTB50 and UAV123.The experimental results show that compared with the feature fusion network,the feature fusion network combined with the attention mechanism can further improve the tracking performance.This thesis proposes a robust adaptive visual tracking algorithm by combining feature fusion and attention mechanism.The experimental results show that the proposed method can significantly improve the performance of the tracking algorithm.On the basis of the research of this thesis,the future work is prospected,including the use of optical flow information to select the target search area,and the combination of semantic information and reinforcement learning and other advanced technologies to improve the performance of the tracking algorithm.