Application Research On Real-time Target Tracking Based On Deep Learning

Target tracking,as one of the key points of computer vision research,has extremely important application value in criminal investigation,intelligent monitoring and intelligent transportation.At present,target tracking faces many challenges,such as complex and changeable environment,non-rigid deformation,occlusion,scale transformation and so on.In the field of target tracking,feature extraction is the most important step.The quality of extracted features directly affects the accuracy of target tracking algorithm,and is also the guarantee of tracking robustness.How to extract robust and abundant features is a problem that needs to be considered in this paper.At present,the main target tracking algorithms are traditional target tracking methods and target tracking algorithms based on deep learning.Traditional target tracking methods perform well in real-time tracking,but the accuracy and robustness are limited in different application environments because of the limited feature extraction ability.The feature extraction ability of target tracking algorithm based on deep learning is very strong,and the biggest challenge is real-time.In view of this,this paper makes an in-depth study of the target tracking network framework based on in-depth learning,weighs the accuracy and real-time,improves the structure of the tracking network,and achieves the goal of accurate real-time target tracking.This paper contains the following research contents:(1)Analyzing the existing target tracking algorithms based on deep learning,the main problem is the real-time performance.In-depth study of Muti-Domain Networks shows that the algorithm has the advantage of high accuracy,but low real-time performance.This paper analyses and improves its structure and tracking strategy in order to prompt you to track speed,and proposes a simple tracking framework,Fast Deeplearning Tracking Networks.Compared with the original algorithm,it solves the problem of low speed.The speed is increased by about 10 times,reaching 20.5 FPS on PC.(2)Although FDTNet basically achieves real-time tracking on PC,the speed improvement on Jetson AGX Xavier has reached a bottleneck,and sacrifices stability to improve real-time performance.The universality of the algorithm is not good.In order to meet the needs of Engineering applications,it is necessary to achieve basic real-time on a small embedded platform,and to track the target stably and accurately.The application of cross filtering in tracking framework is analyzed,and the existing algorithm Fully-Convolutional Siamese Networks is improved to further improve the tracking speed and accuracy.An improved algorithm,Muti-Features Siamese Networks,is proposed to introduce feature fusion and void convolution into the convolution layer,and to improve the tracking speed and accuracy.On PC,different test data sets are tested,and the minimum speed is 40 FPS,which meets the real-time requirements.(3)The algorithm simulation test on PC proves that Muti-Features Siamese Networks algorithm has excellent real-time performance and accuracy.It can be loaded into Jetson AGX Xavier hardware platform and applied in engineering projects.Engineering application needs to train the tracking network framework on the server,then load it into the small hardware platform Jetson AGX Xavier to test the actual video tracking performance.On Xavier,the speed of Siam-MF algorithm can reach 26.5 FPS,which is 2.3 times lower than that of PC,and basically meets the speed requirements of vehicle system.(4)The advantages and disadvantages of this algorithm in application are analyzed,and future improvements are prospected.This paper summarizes the development trend of target tracking algorithm based on deep learning in the future,the development requirement of hardware and the improvement possibility of software.In conclusion,aiming at the problem of real-time target tracking based on in-depth learning,this paper makes a thorough study on improving the tracking network framework,tracking strategy,convolution structure and feature fusion,and proposes a target tracking algorithm to ensure real-time and accuracy.The proposed algorithm is applied to practical engineering projects to meet the speed requirements and track the target steadily.It has the ability of real-time application.