Research On Visual Tracking Based On Deep Learning

Visual tracking is an important research topic in the field of computer vision.It has been widely used in intelligent security,autopilot,human-computer interaction,automatic navigation of unmanned drone,etc.,and has attracted more and more attention in recent years.The ideal tracking algorithm should have the robustness to the change of the target,which means that it can recognize the target in the state of rotation,deformation,occlusion,etc.,and needs to satisfy the adaptability to the background,which means that the target and the background can be correctly distinguished,especially the background which is similar to the target.For a single tracker,it is difficult to keep robustness and discrimination at the same time when updating the model to adapt to the change of target and to prevent model drift.In order to solve these problems,this paper proposes to design a long-term tracker that can maintain the long-term appearance of the target and a short-term tracker with sufficient discrimination for similar target interference,and to achieve better tracking performance by combining the two.On this basis,this paper designs two effective tracking algorithms based on deep learning model.The specific research contents are as follows:In order to obtain different classification performance,based on the different training samples,this paper proposes a robust target tracking method based on dual classifiers trained with hard samples.In this paper,the training samples are divided into positive samples,negative samples,hard positive samples and hard negative samples.The long-term classifier is trained by using long-term positive samples and negative samples,so that it can remember the long-term appearance of the target.The short-term classifier is updated by using short-term collected hard positive samples and hard negative samples,so that it can overcome the short-term effects of similar target interference and occlusion.Considering that it is not reliable to classify targets by maximum score method when tracking fails,this paper proposes an idea based on density clustering to assist target localization.The effectiveness of the proposed target tracking algorithm is verified by extensive experiments on the common video tracking datasets.In order to further improve the performance and efficiency of the algorithm,this paper further proposes using the siamese region proposal network of off-line video pre-training as the long-term tracker,using the classification network based on feature sharing as the short-term tracker,and proposes an effective strategy to integrate the long-term tracker with the short-term tracker.The proposed target tracking method based on siamese region proposal networks and classified networks is tested in various common video tracking evaluation sets.The results show that its main performance indicators(including overlap rate and robustness)are at an advanced level in the current publicly published similar research results.