Research On Single Target Tracking Algorithm Based On Triplets Neural Network

Visual Object Tracking is one of the most important research topics in computer vision.It has important application value in many fields such as driverless,intelligent security,human-computer interaction and information investigation.Due to the complexity and uncertainty of the target's apparent information in the real environment,there are still many difficulties in the single-target tracking task,such as: occlusion,motion blur and background confusion.In recent years,machine learning techniques are often used in computer vision because they have the ability to use a large amount of training data to improve performance.Unfortunately,it does not benefit from massive offline training of video data,and does not take full advantage of the background information of the picture,which resulted in poor robustness and slow computing speed.In order to make full use of the massive video data and make full use of the background information of the picture,this paper transforms the single-target tracking model into the similarity measurement learning model.A new triplets neural network is proposed to implement the similarity measure learning model.During the training phase,the model learns the similarity measure between the targets in the video sequence,the similarity measure between the backgrounds,and the spatial context between the target and the background.In the test phase,the model makes use of the matching relationship between the targets,and the relationship between background and target.The tracker predicts the exact position and size of the target in the current frame,after knowing the information of the target in the previous frame.So loop back and forth,real-time prediction of the location and size of the target achieves the purpose of tracking.In order to enhance the robustness and distinguish ability of the target and background models,this paper analyzes the characteristics of the deep convolution feature for the single-target tracking task.It is found that the high-level convolution features are more semantic and have a more robust representation of the target,which can cope with the significant changes of the target's appearance.The low-level convolution features are more capable of characterizing the image texture and the details of the boundary which enhance the distinguish ability of background.Through the comprehensive use of multilayer convolution features,and on-line learning of the weight of each convolution feature,the objective and background appearances are accurately modeled.In order to adapt to the complicated tracking scenarios and the apparent changes of the target in reality,this paper proposes a method of adaptively measuring the confidence of the tracking results.The target template is updated flexibly according to the confidence of the tracking results to adapt to the change of target outlook,which can reduces the tracking drift caused by unreliable tracking results.Based on the proposed triplets network model,the multi-layer convolution feature fusion method and the adaptive confidence feedback template updating method,an effective single-target tracking algorithm is designed and verified on two public datasets.