Research On Technology Of Visual Object Tracking Based On Unconstrained Environments

Visual object tracking is one of the most important and challenging tasks in computer vision research.It is widely used in many fields such as intelligent video surveillance,unmanned driving,motion analysis,service robots,and so on.Single object tracking is usually used in the current robot application background,so the main task of this thesis is single object tracking.Since the accuracy of visual object tracking will be affected by unconstrain environments,such as illumination variation,motion blur,occlusion,scale variation,and so on.Therefore,research on visual object tracking in this thesis has important significance and broad application scenarios in unconstrained environments.First of all,this thesis expounds the current research status of visual object tracking technologies at home and abroad,it summarizes the key technologies of visual object tracking and the current shortcomings of visual object tracking algorithms,it also clarifies that the main research content of this thesis is feature extraction and tracking model optimization.Then,the overall scheme of visual object tracking system in unconstrained environments is designed,and the existing tracking object state initialization technology is analyzed.Aiming at the problem that the single feature extraction algorithm cannot effectively extract object feature when the object is interfered by unconstrained environments such as illumination variation,motion blur,occlusion,scale variation and so on.This thesis proposes an independent component analysis with reference(ICA-R).Firstly,convolutional neural network with medium structure(VGG-m)network is used to obtain deep features.Then,traditional manual features are combined with deep features in combination with ICA-R.And affine transformation is used to add multiple perspectives to expand the three-dimensional information of the original image and obtain more discriminative features.The experimental results show that the method using the fusing of deep features and traditional mutual features can distinguish between the background and the object.At the same time,it also obtains good tracking precision,tracking success rate and tracking accuracy.In order to solve the problem that Kernel Correlation Filter(KCF)tracking model is likely to cause tracking failure under unconstrained environments such as illumination variation,motion blur,occlusion,scale variation and so on.This thesis proposes an improved KCF and model adaptive update algorithm.First,this thesis proposes a multi-weight KCF,and combines the size of the response values in the previous frame of the object to quickly estimate the scale of object at the current frame.In the model update phase,an adaptive updating strategy based on the center shift Euclidean distance of image patch is proposed to reduce the unnecessary calculation in model training so as to speed up the tracking speed.The experimental results show that the tracking precision,tracking success rate and tracking accuracy of our algorithm are improved,and it shows better robustness in unconstrained environments such as illumination variation,motion blur,occlusion,scale variation and so on.Finally,a service robot control system based on visual object tracking is constructed in this thesis,and the results of visual object tracking are converted into corresponding control instructions to complete the implementation on the intelligent service robot.The experimental results show that the visual object tracking algorithm in this thesis has good accuracy and can effectively control the motion of the service robot.