Research Of Single Object Tracking Based On Memory Network

Single object tracking is a fundamental branch in visual object tracking(VOT).A number of single object tracking models based on deep learning have emerged,where models based on Siamese neural network has become the most influential for they achieve a balance between tracking performance and real-time tracking.The parameters in these models are calculated based on the initial appearance of the object,and remain fixed during tracking.However,the object appearance in subsequent frames of the video may change significantly compared to the initial object appearance.If the tracking model does not update the parameters,it will not be able to learn the changes of object appearance,which will severely affect the tracking performance.Thus,it is necessary to update the model parameters online during the tracking process to learn the changes of object appearance.The historical tracking results contain the appearance information of the object,which can be utilized as samples for online-updating the tracking model.This thesis studies how to use historical tracking results to update tracking model parameters online,and proposes the following two improvements:(1)Single object tracking model based on position attention mechanism and memory network.Our baseline,Mem Track,saves historical tracking results in the memory through write controller,and reads tracking results from the memory through read controller to update the tracking model online.However,Mem Track can generate chaotic write/read weights and has difficulties in convergence during training.To tackle the problems,the thesis proposes a memory reading controller based on position attention and a memory writing controller based on Gaussian Mixed Model.With the improved controllers,the generation of read/write weights is highly related to the characteristics of the current frame.Besides,by separating the write controller from the read controller,we simplify the training procedure.Finally,a single-object tracking model Attn-Mem Track is designed on the basis of the improved memory read and write controllers.Compared with Mem Track on OTB100,the tracking precision is increased by 1.4%,and the tracking accuracy is increased by 1.9%.(2)Single object tracking model based on Few-Shot learning and memory network.As the time interval between the current frame and the initial frame increases,the object in the historical tracking results cannot be guaranteed to be in the center area,named as tracking misalignment.Background noises would be introduced into the memory,leading to a deterioration of feature quality stored in memory and finally affecting Attn-Mem Track's ability in learning the changes of object appearances by updating the model with information read from memory.To solve the problem,this thesis proposes feature reconstruction module based on encoder-decoder to reconstruct the feature of tracking results.The encoder uses dilated convolution to encode global information of tracking result into the central area of output,and the decoder applies the output to reconstruct the object consistent to real object in the central area under the supervision of Perceptual Loss and Triplet Loss.Besides,we train the module with Episodic Training Mechanism mentioned in Few-Shot learning to tackle the problems of few training samples,low convergence speed,and unable to reconstruct in real time.Finally,by integrating the reconstruction module with AttnMem Track,a tracking model Attn-Mem Track-FSR is proposed.The model is proved to be robust and has gained a 4.7% increment in tracking precision and a 2.9% increment in tracking accuracy compared to Attn-Mem Track on OTB100.