Research On Visual Target Tracking Algorithm Based On Siamese Neural Network

Visual tracking is widely used in video analysis,intelligent transportation,virtual reality and other fields.As an important part of artificial intelligence technology,it is the research focus and hotspot of scholars at home and abroad.With the research and development of visual tracking in the past few years,it has gradually encountered a bottleneck,but the rise of deep learning technology has injected new vitality into the research of visual tracking,which has achieved an amazing performance jump and promoted the field to a new stage.However,there are still many problems in the process of tracking the target accurately and robustly.This paper is based on the siamese network,two visual object tracking algorithms are proposed,and a single-target real-time tracking system is designed,the main contents of which are as follows:(1)In order to solve the problems that most siamese network tracking algorithms have weak feature extraction ability of backbone network and the template does not adapt to the change of target.In this paper,the siamese network tracking algorithm based on deformable convolution(DCSiam)is proposed on the basis of the Siam FC algorithm..First,the deformable convolution module learns the adaptive offset of multilayer feature data in different directions to increase the effective receptive field in the convolution process.Then,the final response map is obtained by multi-layer deformable cross-correlation fusion to enhance the deep semantic feature extraction ability of the backbone network.Finally,a high confidence template online update strategy is adopted.The peak sidelobe ratio and maximum value of the response map are calculated every fixed frame as the update basis,and the weighted features are fused to update the template.Evaluate the performance of the algorithm using multiple public datasets.On the OTB2015 dataset,the overall accuracy and success indicators of the DCSiam algorithm are improved by 9.5% and 7.5% respectively from the baseline,which well realizes the accuracy and stability of target tracking in complex scenarios,and verifies the advanced nature of the proposed algorithm.(2)For most siamese network tracking algorithms,they only focus on the similarity between the template and the search area,and ignore the importance of historical frames in the tracking process to the target gradient process and the complexity of traditional bounding box prediction.Based on the Siam FC algorithm,we propose a siamese network tracing algorithm based on spatiotemporal attention(TSASiam).Firstly,the non-local operation of the space-time attention module is embedded in the backbone network,and the network model is assisted to learn the most discriminative historical frames and intraframe regions in the video sequence.Subsequently,the pixel-by-pixel classification regression method is adopted to distinguish the foreground and background of the pixels in the response map,and the regression is regarded as the relative bounding box,and the distance between the pixel point and the prediction bounding box is calculated,and the low-quality bounding box is suppressed by the joint centrality branch,which reduces the prediction complexity and improves the target positioning accuracy.Evaluate the performance of the algorithm using multiple public datasets.Experimental results show that on the OTB2015 dataset,the overall accuracy and success indicators of TSASiam algorithm are improved by 4.35% and 6.00% respectively compared with the baseline,which better achieves the accuracy and stability of target tracking in complex scenarios,and verifies the effectiveness of the proposed algorithm.(3)Based on the twin network target tracking algorithm based on deformable convolution proposed in this paper,a single target real-time target tracking system is developed based on the flash web application framework.The user enters the specified website to enter the login interface,enters the account and password to log in to the back-end interface of the system,and then uses the button and mouse to interact with the server to select and continuously track the video sequence uploaded by the user or the real-time picture of the camera.The client pulls the rectangle to provide the target initial location information,the server invokes the algorithm model,completes the initialization work using the information provided by the client,provides the tracking and positioning service,and returns the tracking and positioning data and picture to the user interface,and then logs each tracking service.Finally,the performance test of multiple modules in the system is carried out,and the successful operation effect is demonstrated,which verifies the reliability of the single-target tracking system designed in this paper and the engineering significance of the algorithm.