Research On Target Tracking Algorithm Based On Network Modulation

The visual object tracking algorithm requires that the tracker can overcome the illumination,deformation and other factors to track the target in the subsequent frames,while the target of the first frame is given.With the wide applications of deep learning algorithm in the filed of tracking,the tracking algorithm has made great process.But in practice,the tracker will also become less robust in the influence of many factors.This thesis proposes the tracking algorithms based on network modulation.The main works are as follows.(1)Aiming at the problem that the high-order information can not be used to model the target appearance robustly,which will make the object drift under the matching idea in tracking.The visual tracking algorithm based on second-order network modulation is proposed.The architecture is divided into two parts,one other is offline training and the other is online update.Offline training learns the modulation vector through the initial frame,and then fuses it into the test frame through the modulation idea.Finally,offline training predicts the intersection over union of gaussian bounding boxes and the labels.At the same time,the tracker combines the deformable convolution block,pixel-wisely Global Contextual Attention and second-order pooling module to make the feature of target more robust.The part of the online update is to learn a classifier to distinguish the foreground and background by the way of exponentially weighted averges in features.Finally,ten bounding boxes are generated according to the rough position and sent to the offline training network.The predicted target box can obtained by maximizing the intersection over union.Through the evaluations of OTB100,VOT2018 and the GOT-10 k datasets,the proposed tracker has greater competiveness with more advance tracking algorithms.At the same time,the speed of this tracker can reach 27 fps.(2)Aiming at the scale change caused by the fast motion of the target,the thesis proposes an object tracking algorithm based on two stream network modulation.In the part of the offline training network,instead of the idea of using two images to match,this paper adopt short video stream as input.And then the 3D convolution is used to learn the motion information of the target.It can been thought to learn the association of time.At the same time,2D convolution is used to learn the appearance of the last frame of the short video.It can been thought to learn the association of space.Finally,the spatial and temporal features are combined to predict the target bounding box by deep bounding box regression network.Instead of the fixed search area,this thesis adopts the strategy that the search area can been adjusted adaptively according the speed of the target.A large number of the experiments show that the proposed tracking algorithm can make full use of the spatial and temporal information and has obvious improvement on OTB100,VOT2018,VOT2019 and GOT-10 k datasets.The final speed of the tracker can reach 12 fps.