Research On Single Object Tracking Based On Deep Learning

Today,in the development of the construction of intelligent city,intelligent video surveillance has become an important part of urban management.The surveillance cameras in the city are playing a vital role in the field of security.Massive video data are stored in a large number of hard drives,so how to fully tap the valuable information has become the focus of the researchers from all over the world.Among them,the target tracking technology for video continue to develop with the increase in the amount of data,and it has become the key to intelligent video surveillance technology.At present,the difficulty lies in the real life scene is more complex and more interference,the computer hardware and tracking algorithm requirements higher real-time quality.Therefore,the study of whether a given scene for any given target to track stably can become the key to breakthrough.The most important part of the traditional single-target tracking model is feature extraction.The characterization ability of the tracking target appearance model determines the tracking accuracy of the tracking algorithm in generalization performance.In this paper,based on the study of traditional color and texture feature for tracking,this paper makes a deep comparative study on the deep convolution neural network with excellent description ability.In order to be able to guarantee tracking accuracy and real-time at the same time,an approach to training an end-to-end offline depth learning model for online tracking is proposed.At the same time,for the given single target in the tracking process of deformation,partial occlusion and loss,etc.,the tracker and the traditional detection framework integration,used to ensure the robustness of tracking.And gives a confidence index to determine the accuracy of tracking,with the target model to achieve self-calibration tracker,and in the long-term tracking process achieve re-detection.The tracking algorithm proposed in this paper is experimentally tested,and combines the advantages of classical tracking algorithm and deep learning algorithm in both accuracy and real-time,and shows a better ability to track.