Convolutional Neural Network For Real-Time Object Tracking With Mutual Learning

Video object tracking algorithm is a very practical algorithm.At present,the object tracking algorithms based on convolutional neural network have shown favorable performance.However,most convolutional neural networks are designed for image classification tasks and are generally time-consuming.Therefore,the object tracking algorithm based on convolutional neural network can not satisfy the real-time requirements in practical applications.Although some object tracking algorithms based on convolutional neural network abandon model updating in online tracking phase to achieve real-time tracking,these algorithms generally have unsatisfying performance.The existing tracking algorithms often fail to take account of both tracking accuracy and real-time performance.Quite a few algorithms ignore real-time performance.This work studies these problems.The main work and innovations are as follows:(1)By combining the respective advantages of correlation filter framework and siamese network framework,a new fully-convolutional neural network tracking framework is designed in this work.Through the analysis of the correlation filter framework and the siamese network framework,this work summarizes the common points of the two algorithm frameworks in the location principle.This work also analyses the main differences between the two frameworks,and discusses the main impact of these differences on the speed and accuracy of tracking algorithms.According to the similarity between the two frameworks,this work designs a tracking algorithm framework based on fully-convolutional neural networks,which incorporates model updating and network training into the tracking algorithm.(2)In this work,a tracking algorithm based on a channel pruning network is constructed.Because of the use of lightweight channel pruning networks,the speed of algorithm has been speeded up.Through the reasonable design of network structure,the accuracy of tracking is guaranteed.At the same time,considering the speed of the algorithm,the model updating mechanism is simplified and a faster scale estimation method is adopted.Finally,the deep mutual learning method is adopted in network training,which further improves the performance of network and tracking accuracy.(3)Through a large number of experiments on three mainstream tracking algorithm evaluation datasets,the rationality of the proposed tracking algorithm and the effectiveness of the deep mutual learning method are verified.The experimental results show that the proposed tracking algorithm can track objects at 60 FPS.The precision score and the success score on OTB2015 are 0.887 and 0.640 respectively,and the EAO score on VOT2017 real-time experiment is 0.261.