Research On Object Tracking Based On Convolutional Neural Network

The object visual tracking algorithm has received extensive attention from researchers due to its great social value.Object visual tracking plays an important role in many areas such as surveillance,transportation,and military.Although the object tracking has been studied for several decades,there are still many problems such as object occlusion,lighting changes,and posture changes in the tracking process.Obtaining a robust tracking algorithm is still worth studying.The traditional tracking algorithm uses manually designed features or a simple combination of features to model the object.The lack of expression of the features makes it difficult to further improve the tracking accuracy.In order to solve this problem,researchers have used convolutional neural networks for object tracking.However,due to the large amount of computation of convolutional neural networks,the real-time tracking performance is generally low.This article starts from the two aspects of accuracy and real-time,and improves the speed of tracking under the premise of ensuring accuracy.From the two aspects of network training methods and neural network models,a new object tracking algorithm for convolutional neural networks based on single extraction features is proposed.First of all,we select a deep neural network model with powerful feature extraction capabilities,integrates the idea of migration learning,and adopts the “offline training +online fine-tuning” model.At the same time,in order to better fit the object tracking problem,fine-tuning training is performed using the referenced video sequence set,so that the model learns the common problem of object tracking.Secondly,online tracking does not directly extract positive and negative samples from the input frame.Instead,the entire image is directly input into the convolutional neural network at one time for forward calculation,and features are extracted once in a single time,according to the position invariance of the convolutional neural network,and the generated features.The corresponding position in the figure finds the positive and negative samples,reduces the repeated calculation of similar regions of the positive and negative samples,reduces the amount of calculation of the algorithm and thus acceleratesthe tracking speed.In order to improve the accuracy,a border regression model is added.At the beginning of tracking,the frame regression model is trained using the information of the first frame,so that the prediction frame is closer to the real frame.Two different update methods includinng long-term update and short-term update are adopted to deal with the drift in the tracking problem effectively.Finally,compared with other algorithms in the OTB sequence set,the experimental results show that this algorithm can accurately deal with the problems of deformation,occlusion and other problems in the tracking process,can significantly improve the tracking accuracy,and have a certain effect of speeding up the tracking speed.