Research And Application Of Visual Detection And Tracking Algorithm

Visual object tracking is a significant yet hot research topic in the field of digital image processing,artificial intelligence and so on,and it has been widely applied in intelligent transport,human computer interaction,visual surveillance,just to name a few.It remains a challenging problem to design a robust and efficient visual tracking due to the object occlusion,scale variation,illumination variation,background clutter,and so on.To cope with these challenges,in this thesis,we concentrate on how to design an efficient motion and appearance model for visual tracking,and some novel tracking methods have been presented.1.We proposed a graph-structured multi-task sparsity model for visual tracking.We model the particle correlation via the relational graph,and propose a novel graph-guided sparse learning model to incorporate the topological constraints of relational graph into multi-task framework.By capturing this underlying relationship,our proposed method encourages related particles in accordance with this graph to be reconstructed by common templates.Based on this,a novel l1-graph structured sparsity framework is conducted to obtain the sparse coefficient.Meanwhile,to preserve spatial and discriminative structure,we deploy an effective observation model which conducts sparse coding on local image patches and augments multiple background information to the Bayesian sequential inference framework.Extensive experimental results on the large scale tracking benchmark verify that our method outperforms the existing trackers.2.We proposed a visual tracking algorithm based on multi-feature and object proposal.Typical correlation filter trackers cannot handle the scale variation problem adaptively.To cope with this challenge,we integrate a class-agnostic object proposal method into correlation filter tracker.According to the target location of feature fusion,we deploy the structured edge detection.By the means of training least squares method classifier of kernel function,and we design an adaptive strategy to compare the maximum value of the object proposal response with the general correlation filter result.Thus,this method can change the size of the tracking frame according to the size of the target and improve the tracking accuracy.Extensive experimental results on the large scale tracking benchmark verify that our method can perform well in terms of the tracking robustness and efficiency.3.We proposed a visual tracking algorithm based on deep feature and objectness proposal.Within the framework of kernel correlation filter tracking method,we apply a pre-trained convolutional neural network and interpret the hierarchies of convolutional layers as linear counterpart of an image representation from coarse to fine.Meanwhile,based on the deep feature expression,this paper integrates the objectness proposal method,which is widely used in object detection area.An effective and dynamical proposal rejection strategy is also adopted to make the proposals more useful.The experimental results on a large-scale benchmark dataset show that this proposed algorithm is more accurate and robust in coping with scale changes and motion blur.