Target Tracking Algorithm Based On AOG And FlowNet2.0 Network

In the field of human and animal vision,target tracking is an inherent ability to learn visual concepts,and it is also an important task in computer vision.Target tracking,especially long-term tracking,is a problem.It not only needs to deal with changes in tracking targets,including appearance changes,structural changes,scale changes,occlusion,etc.but also needs to solve the complexity of the scene,such as camera movement,background clutter distractions,lighting changes,frame cropping,etc.In And-Or graph tracker,in order to solve the problem that the tracking target changes during the tracking process,a hierarchical combination And-Or graph representation is used to simultaneously track,learn and parse the unknown targets in the video sequence,and combines detection and tracking to perform more accurate tracking.The tracking performance of And-Or graph tracker is better than most existing target tracking algorithms,but it also has some problems,such as prone to tracking errors when similar targets appear in region of interest,inaccurate tracking when tracking objects have large displacement,slow tracking speed,etc.Aiming at the problems existing in And-Or graph tracker,this thesis proposes a target tracking algorithm based on AOG and Flow Net2.0 network.The work of this thesis is mainly from the following aspects:(1)In this thesis,in order to improve the accuracy of the tracking algorithm,the Flow Net2.0 network will be used to replace the Lucas-Kanade optical flow method for target tracking.Using optical flow method to track the target,Lucas-Kanade optical flow method and Flow Net2.0 network have better tracking effect when tracking object displacement is small,but Flow Net2.0 network has better tracking effect than Lucas-Kanade optical flow method when tracking target displacement is large.Therefore,using Flow Net2.0 network instead of LK optical flow method can improve tracking algorithm accuracy in large displacement.(2)On the basis of AOG and Flow Net2.0 networks,appropriate selection rules for tracking results are formulated.In the process of selecting tracking results from optical flow prediction results and AOG detection results,different selection rules will produce different tracking effects.The essence of selection rules is to select different candidate targets as tracking results under different circumstances.When similar targets appear,the selection rules of And-Or graph tracker are prone to tracking errors.Therefore,the selection rules of tracking results are re-established in this thesis,which makes the tracking effect better.(3)The tracking speed of the tracking algorithm is improved.The process of relearning the object AOG has a great influence on the tracking speed,so it is important to choose the timing of relearning the object AOG.In And-Or graph tracker,the moment of relearning the object AOG is called the critical moment.Zhu chooses to relearn the object AOG when the frames with invalid tracking are accumulated to a certain number,which would cause unnecessary relearning.Therefore,in this thesis,for the current object AOG,when continuous multi-frame tracking is invalid,the object AOG can be relearned,reducing unnecessary relearning process and speeding up the tracking speed.Through the experimental simulation of the target tracking algorithm based on AOG and Flow Net2.0 network,the results show that the algorithm in this thesis can track more accurately when the tracking target has a large displacement and a similar target appears in region of interest;On OTB-2013 data set,compared with And-Or graph tracker,the tracking success rate of this algorithm in this thesis is improved by 2.8%,the tracking accuracy rate is improved by 3.1%,and the tracking speed is improved by 3.2%.In addition,the performance of this algorithm is compared with the kernel correlation filter algorithm,full convolution Siamese network and tracking-learning-detection algorithm,the tracking accuracy and precision of this algorithm are obviously better than other tracking algorithms,and the tracking speed has certain competitiveness.