Visual Target Tracking Based On Deep Learning

As one of the basic directions in the field of computer vision,target tracking has made great progress in recent ten years.However,target tracking technology has not reached a satisfactorily high-level performance.There is still much room for the improvement of tracking accuracy,robustness and tracking speed.In addition,researchers have not put enough effort on long-term tracking.In recent years,the rapid rise of deep learning makes people see the new dawn of computer vision and artificial intelligence.Target detection and recognition have made great progress through the method of deep learning.However,the application of deep learning technology in target tracking has started relatively late.So there is no representative work in the field of target tracking yet.The title of this work is visual target tracking based on deep learning.Aiming at the problems of low accuracy,poor robustness and short tracking time of existing target tracking methods,this research mainly attempts to improve the comprehensive performance of target tracking from short-term tracking and long-term tracking by using deep learning and provide new ideas for target tracking.The main research contents of this paper are generalized as follows:This study proposes an online convolutional neural network structure for acquiring convolution kernels through unsupervised learning,Which is used as a feature extractor.Besides,kernelized correlation filter(KCF)was combined to the framework to search and locate the target.In addition,this work proposes a new type of update scheme that combines the update of convolution kernel and the update of correlation filters to improve the robustness of target tracking.At last,the algorithm was experimented based on OTB2013.This study designs a network structure based on the large pre-trained off-line convolution neural network model for target tracking.This network abandons the traditional modular design pattern and transforms the target tracking into an end-to-end deep regression network which can be used to predict the bounding box of the target.This study proposes a data enhancement method based on the OTB dataset and build training dataset by this method.At last,the trained network is evaluated in VOT2014.This study proposed a new concept of tracking by detection which is based on a highspeed,high-precision regression deep learning algorithms.Besides,a new framework for long-term tracking is established.In addition,an innovative motion estimation module and Gaussian weighting mechanism are proposed to ensure the continuity of tracking.This study also proposed an improved one-click initialization method to make tracking initialization easier and more accurate.It is worth mentioning that this study also built a long-term tracking evaluation dataset that contains the most common long-term tracking challenges.At last,this study designed a complete target tracking system including hardware design and software design.The performance of proposed algorithms was verified and feasibility was analyzed.The final results show that the target tracking based on convolutional neural network and correlation filter can not complete the tracking task due to the poor real-time performance,while the target tracking based on the offline depth network and the long-term target tracking based on the deep learning regression detection show excellent overall performance.All the experimental demonstrates demonstrate that deep learning can achieve higher tracking accuracy better tracking robustness compared with the traditional methods.In addition,deep learning methods are able to achieve higher tracking speed based on high-speed calculation of GPUs.