Research On Object Tracking Method Based On Multi-domain Learning Network

Target tracking has been the research interest of many researchers in the field of computer vision(CV).Target tracking refers to the estimation of change in position and scale of a given target in a video sequence based on its initial information,whose output will be the input of higher-level tasks.There are many blocks in visual tracking,firstly,the available training data and information about the given target is very limited,and secondly,the tracking model faces many challenges such as occlusion and blurring that exist in reality.In target tracking,target itself together with its background changes are reasons that affect the tracking results.Currently,target tracking technology has been promoted to be used in several real application scenarios such as traffic monitoring,criminal tracking,and video beautification.In summary,it is very important to study the target tracking algorithms for practical applications.In this paper,we launch an in-depth study of the mainstream deep learning target tracking algorithms in recent years to improve the accuracy and speed of target tracking,and optimize the target tracking algorithm based on MDNet target tracking algorithm.The main research work in this paper is as follows:1.A target tracking improvement algorithm combining attention mechanism is proposed for the problem that the feature discrimination ability of the underlying convolutional neural extraction of MDNet algorithm is insufficient.Firstly,the MDNet algorithm is analyzed,and the results show that if the background in the tracking process is complex,the tracking results of the MDNet algorithm are less accurate,mainly due to the insufficient attention to the important features of the target.In view of this,this paper introduces a hybrid attention mechanism containing two attention mechanism modules: channel attention has better learning ability for the target deformation to be tracked in the neural network,while spatial attention can calculate the weights of spatial locations on the feature map and increase the weights of relatively more important features' spatial locations,thus making feature extraction more effective.The attention mechanism is introduced to improve the backbone network model in the original model algorithm,and a multi-domain learning network target tracking algorithm incorporating the attention mechanism is proposed,which can enhance the discriminability of the extracted features and thus improve the discriminative ability of the network model for targets.2.To address the problem that the IOU of multi-domain learning network cannot accurately portray the relationship between the generated frame and the target frame,a goal tracking algorithm combining multiple loss function optimization is proposed.First,GIOU(Generalized-IOU)is introduced to re-measure the problem of non-overlap between the generated frames and the real frames in the target tracking algorithm.In the case of non-intersection,GIOU can better reflect the alignment of the real frame and the generated frame.Secondly,DIOU(Distance-IOU)is introduced to solve the problem of moving direction of the generated frame when the target frame does not overlap by strengthening the measure of the distance between the center point of the generated frame and the target frame.Meanwhile,CIOU(Complete-IOU)is introduced to solve the aspect ratio consistency problem by considering the difference between the aspect ratio of the generated frame and the target frame.In this paper,GIOU,DIOU,and CIOU metrics are introduced to improve the acquisition level of the generated frames and thus improve the performance of the algorithm.