Research On Real-time Visual Tracking Method Based On Neural Network

Object tracking is a challenging research direction in the field of computer vision.In recent years,many excellent tracking algorithms have been proposed and breakthroughs have been made.However,due to the complex and changeable application scenarios,including background clutter,deformation,fast motion,illumination changes,occlusion,scale changes and other influencing factors.The research and implementation of object tracking brings huge challenges.This article first analyzes and studies the existing classic target tracking algorithms and methods.The method based on correlation filtering has very strong real-time tracking due to its extremely high computational efficiency,but the tracking accuracy is poor due to insufficient and unspecific expression of target features.Based on the deep learning method,with the powerful feature representation ability of the convolutional neural network,the accuracy and robustness of target tracking have been greatly improved.However,because the neural network contains many parameters and operations,it takes a lot of time and Space,resulting in poor timeliness of target tracking.In order to solve the problem that the target tracking accuracy and tracking speed are difficult to balance.The main research contents of this paper are as follows:(1)The traditional target tracking algorithm based on convolutional neural network is improved,and the tree structure is used to dynamically update the CNN.Through the joint work of multiple convolutional neural networks in the tree structure,the state recognition of the target is completed.At the same time,the CNN in the tree is constantly updated to enhance the performance of multi-modal processing of the appearance of objects.This is of great help to the tracking of high-speed moving objects.Experiments have proved that the proposed algorithm does not significantly increase the computing performance requirements,and at the same time can support faster identification and tracking with higher accuracy.(2)A shallow network with additional judgment mechanism is constructed,which can automatically select feature matching based on the number and frequency of interferences.When a frame of image contains only the target or a small number of analogues,selecting low-level features can easily mark the target;otherwise,using high-level semantic features.In this way,accurate positioning can be reduced,and the computational complexity can be reduced,and the tracking speed can be improved.In the video sequence,the position of the object in the current frame does not shift much from the position of the previous frame.Through the above features,it is easy to establish the connection between adjacent frames and predict the position of the target in the next frame quickly and accurately.(3)By analyzing the characteristics of Siamese network,a new target tracking method is proposed.For each branch of the Siamese network,an Alex Net shallow network and an improved Res Net deep network are constructed at the same time for feature extraction.At the same time,further classification and regression are carried out,and learning is carried out in an end-to-end manner at the same time.A new tracking strategy is proposed to perform multi-feature fusion on the extracted features to enhance the discrimination and improve the accuracy of target tracking.The search strategy from local to global is adopted to reduce computational complexity and reduce the waste of time and space resources.The algorithm was tested on the target tracking standard data set(OTB,VOT).Through the quantitative and qualitative analysis and comparison with some existing trackers.Experimental results show that the proposed algorithm can improve the tracking accuracy while ensuring the tracking speed.At the same time,the target can be tracked robustly under the influence of deformation,occlusion,and background clutter.