Research On Single Target Tracking Algorithm Based On Deep Learning In Complex Scenes

Target tracking technology is one of the important research directions in the field of computer vision.It is widely used in transportation,surveillance,military and other fields,and has extremely important practical value and research significance.The target tracking task refers to predicting the position coordinates of the target in the acquired frame according to the position of the target in the initial video frame,and marking it with a bounding box.In recent years,with the advancement of science and technology,the target tracking technology has been greatly developed,and a large number of excellent target tracking algorithms have emerged.However,the difficulty of motion blur and low resolution in the target tracking field is still not well solved,so complex scenes Achieving robust real-time tracking of targets is still a challenging topic.The existing target tracking algorithms are mainly divided into traditional target tracking algorithms and deep learning target tracking algorithms.Although traditional target tracking performs well in a specific simple scenario,the tracking effect is poor in complex scenarios and cannot meet the actual application requirements.The deep learning target tracking algorithm can well cope with various interference factors in complex scenes,but its feature extraction network structure is complex and computationally intensive,resulting in poor real-time performance.In view of this,this thesis conducts an in-depth study on the target tracking algorithm based on deep learning,weighs the accuracy and real-time,selects the appropriate tracking network framework,and improves the requirements of large robust real-time tracking.The main research contents of this paper are as follows:1.This article studies the basic definition of target tracking,the tracking process,and the related difficulties of the tracking process.It analyzes and introduces the role of each module of the target tracking basic framework in target tracking and the key technologies involved.The structural characteristics and working principles of feedforward neural networks and convolutional neural networks are studied,which lays a theoretical foundation for the subsequent chapters.2.In this paper,the existing target tracking algorithm based on deep learning is compared and analyzed,and the accuracy and real-time of the tracking algorithm are weighed according to the actual application requirements.Finally,the target tracking algorithm based on the full convolution twinning network(SiamFC)is selected as the basic tracking network.The framework,and the use of shallow AlexNet as the feature extraction network for SiamFC,leads to the analysis and introduction of the problem of poor target feature representation.For the motion blur and low resolution,SiamFC can not effectively extract the target features,the model drifts,which leads to the problem of tracking failure.This paper improves the original SiamFC and proposes a target tracking algorithm based on conditional confrontation to generate twin networks(CAGNSiamFC).The deblurred network module is generated by embedding the condition in the SiamFC tracking network to deblur the input motion blur and low resolution video frames,so that the feature extraction network can effectively extract the target features and improve the dynamic adjustment capability of the tracking network.Tracking algorithm robustness,tracking of targets under motion blur and low resolution.3.In order to reduce the training time and reduce the training difficulty,this paper adopts the idea of migration learning,and uses the separate training online combination method to train and test CGANSiamFC.During the training,different training data sets are used to optimize the SiamFC tracking network and the CGAN de-fuzzy network module.During the test,the trained SiamFC tracking network and the CGAN de-fuzzy network module are embedded and combined.Finally,the CGANSiamFC proposed in this paper was evaluated using the OTB100 test set,and the evaluation results were compared with the original SiamFC and other tracking algorithms.The experimental results show that the proposed algorithm can accurately deal with the motion blur and low resolution problems compared with the original SiamFC,and significantly improve the tracking accuracy.