Research On Object Tracking Algorithm Based On Deep Learning

Object tracking is one of the important branches of the computer vision.With the development of information technology,it has getting more and more attention and application in the fields of human-computer interaction,intelligent robot,automatic driving,national defense security,video surveillance,smart city,etc.Although visual object tracking technology in the past dozens of years got rapid progress,but due to the complex and changeable tracking environment factors such as object occlusion,scale change,appearance deformation and similar objects interference,it is still a difficult but promising job to meet the real-time,precision and robustness requirements of tracking in multiple application scenarios.Based on the deep learning algorithm model,this paper conducts in-depth exploration and research on the problems that tracking drift or loss under complex conditions such as target deformation,occlusion and interference by similar objects,and proposes several improvement scheme to improve the tracking accuracy and robustness,and meet the real-time tracking requirements in specific application scenarios.The main research work of this paper is as follows:(1)Aiming at the problem of model tracking drift caused by the occlusion and out of view of the target in the long time tracking,an object tracking method which can effectively update parameters online is proposed based on the binary classification network model.First,an improved shrinkage loss function is introduced to solve the problem of positive-negative class imbalance in model online training.Second,a high confidence retention sample pool is designed to retain the valid and highest confidence results of each frame during online tracking,and to replace the lowest confidence retention samples when the pool is full.Finally,when the model detects tracking failure or the continuous tracking frame number reaches a specific threshold,the reserved sample pool is used for online training to update the model,so as to make the model robust and efficient in dealing with long-term tracking.Experiments show that the method has good robustness and strong competitiveness in tracking accuracy and success rate.(2)Aiming at the problem that the Siamese network tracking model is difficult to deal with the complex tracking factors such as target deformation and similar object interference,an fast object tracking method combining residual connection and channel attention mechanism is proposed.First,the shallow structure features and the deep semantic features extracted from the template branch network are effectively fused through residual connections to improve the model's representational ability.Second,the channel attention module is introduced to make the model adaptively weighted to different semantic target feature channels to improve the generalization ability of the model.Finally,a weight mask based on correlation response values is designed and proposed to increase the weight of similar semantic target loss values during offline training,so that the model is enhanced discrimination of similar semantic targets in end-to-end offline learning.Experiments show that the method has good tracking performance and superior real-time performance.(3)Based on the Flask application framework,the designed and implemented tracking algorithm in this paper is implemented in a Web programming system.Users can interact autonomously with the server in the browser page of the program to select multiple target object in the real-time video frame stream captured by multiple cameras,read the positioning data of each target from the back-end algorithm model,and reset the tracking target,etc.The server side monitors the user's behavior in real time,locates and tracks the target selected by the user,and sends back data such as image and location to the front page for display,and generates log files.The functional and non-functional tests verify the stability,compatibility and reliability of the developed system,and the running results of the system demonstrate the real-time performance,robustness and practical significance of the proposed algorithm.