Robust Visual Tracking Under Complex Scenarios

With the development of UAV industry and the progress of artificial intelligence,the application of detection and tracking system based on machine vision platform is becoming more and more common.Visual tracking provides core solutions for machine vision problems which draws the attention of many researchers.Recently,researchers from China and abroad have proposed some novel feature extraction method and model updating strategy based on convolutional neural network.However,the tracking algorithm still fail due to the complex scenarios.Therefore,the accuracy and robustness of the visual tracking algorithm need to be deep research and the combination of the visual tracking algorithm with UAV platform and some other machine vision platform is also an urgent engineering problem.This essay aim at dealing with robust visual tracking under complex scenarios with the references from China and abroad and obtaining more accurate and robust algorithm and visual tracking strategy is applied to the UAV tracking in order to meet the needs of practical application.The main work is as follows:1.Aiming at solving the problems of visual tracking failure caused by occlusion,deformation and similar objects existing in online model updating method and offline networks visual tracking method,a visual tracking algorithm combining online with offline training method is proposed.The proposed algorithm first to introduce the online updating method based on template and histogram.Then the fully-convolutional Siamese networks is illustrated.Finally,the relationship between the disperse measurement and adaptive threshold is employed to determine the switching time between online method and offline method during tracking process.Experimental results show that the proposed algorithm can combine the advantages of online method and offline method effectively to avoid the tracking drift when the target encounters occlusion and similar object which has a higher tracking success rate,and ensures the real-time performance of the tracking algorithm at the same time.2.Visual tracking algorithm based on Siamese networks employ the convolutional feature of deep layer which is lack of the spatial location information.We propose the tracking algorithm based on hierarchical Siamese networks.Firstly,the framework of hierarchical Siamese networks is designed.Then the hierarchical Siamese networks is trained by end-to-end learning to estimate the similarity function between the target and candidate target.Finally,the visual tracking process is completed by locating the position with the highest similarity score as the evaluated target.The experimental results show that the algorithm is well performed in accuracy,because the target features obtained by the hierarchical Siamese network contains both spatial information and semantic information.The proposed algorithm also shows strong robustness when the target suffering occlusion and scale changes.3.The conventional algorithm performs too slow due to the high feature redundancy in feature extraction based on convolutional neural network.When the target is occluded,the visual tracking will drift easily.Due to the absence of a long-term tracking strategy,the long-term tracking is facing with the low success rate.The proposed method employs adaptive dimensionality reduction,adaptive model updating and combination of offline Siamese tracker to ensure the efficiency of the algorithm and the robustness.Firstly,the spatial information and semantic information contained in the convolutional feature are weighed and the adaptive projection matrix is used to obtain the principal component feature of the swallow layer.Then,the peak-to-sidelobe ratio is used to assign weight to adaptive model updates.Finally,combined with the offline Siamese tracker,the offline tracker is activated at the appropriate time through threshold activation to complete the overall tracking process.Experimental results show that the proposed method is more efficient than the conventional tracking algorithm based on convolutional features,and performs strong robustness in the attributes of scale change,occlusion,camera motion,scale change and deformation and has achieved great performance in the visual benchmark.4.Aiming at the problem that the features of convolutional network are not robust enough to represent the appearance of the target when it is suffering affine transformation during tracking,and the bounding box is not affine robust to the deformation.The tracking algorithm based on spatial transformation Siamese networks is proposed.Firstly,We construct the spatial transformer Siamese networks and obtain the affine transformer matrix by localization network,grid generator and sampling module.Then the candidate target which has the highest similarity is obtained by employing the similarity matching method through end-to-end learning.Finally,affine bounding box regression of the proposed tracking is achieved by inverse transformation of affine transformation matrix.Experimental results show that the proposed algorithm has better tracking performance than other tracking algorithm when deformation occurs in the tracking process.Affine boundary box improves the tracking accuracy and makes the tracking algorithm more practical.5.Aiming at the problems of insufficient high-quality and weak diversity of training samples,tracking failure in the long-term tracking process suffering complex scenarios.We propose the tracking algorithm based on adversarial fusion network.The proposed algorithm first to introduce the structure and training method of the adversarial dropout network,then proposes the adversarial spatial transformer network and the fusion method of this two networks,finally employs the adversarial fusion network to generate high-quality proposals to complete the classification of candidates and online update is achieve by the proposals.Experimental results show that the algorithm performs robust to different occlusion,deformation and fast motion.The visual tracking algorithm still has the opportunity to relocate the target when the target re-appear since the algorithm is implemented by proposals.6.Compare the overall performance of the visual tracking algorithm which are proposed in the essay.Through the tracking performance of the proposed algorithm and real-time evaluation,it is concluded that the proposed algorithm has different pros and cons during tracking in complex scenarios and can solve the specific problems.The conclusion also figure out the future researches of the proposed methods.The feasibility and effectiveness of the proposed algorithm in solving more complex scenarios such as UAV tracking task are verified by the comparative experiment of the proposed algorithm in the UAV tracking benchmark.In conclusion,this essay are aim to solving the problems that exist in the visual tracking complex scenarios.The essay does the researches and provides corresponding solutions which achieve the better performance in visual tracking benchmark and UAV tracking benchmark which provides visual tracking algorithm for the practical application.In the end of the essay,the main work is summarized,and the future research is prospected.