Research On Visual Tracking Algorithm Based On Convolution Neural Network

With the rapid development of high-pixel imaging equipment and computer data processing,tremendous progress has been made in the field of computer vision.Visual target tracking is an important branch of computer vision and image processing.It has a wide range of applications in the fields of human-computer interaction,autonomous driving,military,and image target recognition.The emergence of deep learning has solved many complex problems in the field of computer vision,and has achieved good results in target detection,image segmentation and classification.In recent years,many target tracking algorithms based on deep learning methods have been proposed,which has accelerated the development of target tracking and greatly improved tracking accuracy and speed.This article mainly analyzes and studies target tracking algorithms based on convolutional neural networks,and uses its advantages to optimize and improve existing algorithms.The main research contents and innovations are as follows:The article has made a comprehensive review of the research status of target tracking algorithms,understands some of the current challenges in the field of target tracking,and introduces in detail neural networks,target tracking methods based on convolutional features and correlation filters,and the target tracking methods based on end-to-end the network model,found that rich feature expressions and reduced complex calculations will make target tracking performance more robust.(1)In order to be able to fully extract the convolutional features of the target and use its semantic information to better track in more complex scene changes,this paper proposes a robust visual tracking algorithm based on spatial-temporal context hierarchical response fusion.The algorithm uses the layered features of the convolutional neural network to learn the spatial-temporal context correlation filters on the convolutional layer and locates the target position by fusing the response values ??of the filters on the three convolutional layers.In terms of scale estimation,learn a scale discrimination correlation filter to estimate the target scale from the most ideal response value.In order to better solve the tracking failure situation,a re-detection activation discrimination method is proposed to improve the robustness of target tracking.In addition,in order to make the model higher quality,an adaptive model updating method is proposed,the tracking drift caused by noise update is reduced to a certain extent.A large number of experiments have proved that the proposed algorithm has higher accuracy and success rate,and has good performance in complex scenes such as background clutter,occlusion,scale variation,illumination variation and in-plane rotation.(2)In order to better improve the performance of target tracking,balance its tracking accuracy and tracking speed,based on the siamese network framework,this paper proposes a robust adaptive learning with siamese network architecture for visual tracking.The algorithm fuses traditional features and deep convolution features,and a feature adaptive fusion method is proposed to improve the effectiveness of feature expression.In order to further improve the generalization ability of the model,the tracking model of the previous frame is updated by the learned target appearance change factor and background information change factor,and a model update strategy is proposed,and reduces tracking drift in the event of tracking failure,occlusion or background clutter.The algorithm has performed a lot of experiments on public benchmark data sets.The experimental results show that the proposed algorithm has higher accuracy and success rate,and has also greatly improved in speed.The overall performance is ahead of the current mainstream some tracking algorithms.