The Research On Object Tracking Based On Segmented Fine-Grained Regularization And Data Augmentation

Object tracking is a basic task in the field of computer vision,which has been widely used in military security,intelligent monitoring,human-computer interaction and augmented reality,and has achieved a series of important application achievements.The main idea of the object tracking algorithm is to build a model to learn the information of the first frame of the video sequence,and then accurately identify and locate the object in the following consecutive video frames.Siamese network is the current mainstream tracking model,which regards object tracking as a similarity learning problem,learns the similarity between the tracking object and the search area through end-to-end offline training.Also,Siamese network can learn the general relationship between object motion and appearance from a large amount of data labeled with true position and scale,so as to identify and locate objects that have never been seen during training.However,since the Siamese network tracker does not perform model update,its robustness completely depends on the generalization ability of the Siamese network architecture,and drastic changes in the object appearance will lead to a large reduction in accuracy.In order to improve the robustness of Siamese network trackers,many existing studies have introduced such as attention mechanism,interference perception module,online model update or deeper and wider network structure.However,the above methods just focus on the improvement of network structure,but ignore the optimization of network weights.In fact,the generalization ability of the network depends not only on the structure of the network,but also on the weights of the network.Therefore,this paper focuses on the robustness improvement of the object tracking algorithm,effectively combines the regularization method in machine learning and the image data augmentation method in deep learning with the Siamese network tracking algorithm,and verifies the value of the improved algorithm with extensive experiments.The main research contents of this paper are as follows:（1）Most of the siamese network tracking algorithms use L₂ regularization in the training stage,while ignoring the hierarchy and characteristic of the network architecture.As a result,such trackers have poor robustness.With this insight,we propose a segmented fine-grained regularization tracking（SFGRT）algorithm,which divides the regularization of siamese network into three fine-grained levels,namely filter level,channel level and shape level.Then we creatively build a segmented fine-grained regularization model that constructs penalty functions based on group lasso,which combines with different levels of granularity to improve generalization ability and robustness.In addition,aiming at the imbalance of gradient magnitude of each penalty function,our approach constructs a gradient self-balancing optimization function to adaptively optimize the coefficients of each penalty function.Finally,ablation study on VOT2019 show that compared with the baseline algorithm SiamRPN++,our approach achieves relative gains of 7.1%and 1.7%in terms of robustness and expectedaverage-overlap（EAO）metrics,respectively.It means that the robustness of our tracker is significantly enhanced over baseline tracker since the smaller the robustness metrics,the better.Extensive experiments based on VOT2018,VOT2019,UAV123 and LaSOT show that the proposed algorithm has better robustness and tracking performance than related state-ofthe-art methods.（2）The existing information dropping algorithms have the disadvantage of completely erasing the tracking object,which are unfavorable for object tracking.Therefore,a Scale Adaptive Information Dropping Augmentation based Tracking（SAIDAT）algorithm is proposed.Scale adaptive information dropping augmentation approach implements image information dropping by generating a zero mask to occlude image blocks,and adaptively reshape the size and shape of the zero mask according to the scale of the bounding box of the tracking object,thereby avoiding completely erasing the tracking object.Scale adaptive information dropping augmentation results in the effect equivalent to regularization in network training stage,which can enhance the generalization ability of the network and improve the robustness of the tracker.The experimental results obtained on VOT2019 demonstrate the merits of the proposed SAIDAT and its superiority over the state-of-the-art trackers.