Research On Object Tracking Method And Key Technology Based On Sparse Representation

As one of the key technologies in computer vision,object tracking has widely been used in many fields,such as video surveillance,intelligent transportation,human-computer interaction,military,medical treatment and augmented reality.In the past decade,both domestic and foreign researchers have carried out numerous researches on object tracking and have achieved remarkable progress.However,it is still of great theoretical significance and practical value to study a robust and feasible object tracking method due to the complexity of realistic application scenarios(illumination variation,background clutters,occlusion,deformation,scale variation,rotation,fast motion and motion blur).This dissertation reviews the literatures from both domestic and foreign researchers and does a deep research on object tracking method based on sparse representation.The main contributions and innovations of this dissertation are as follows:1.A novel object tracking algorithm with incremental subspace learning constrained sparse representation model is proposed.Sparse representation tracking algorithms can deal with the occlusion problem in object tracking effectively.They use a dynamic template updating strategy in order to handle the target appearance changes during tracking,but the strategy only preserves the information of target object appearances with a previous couple of time instants,thus cannot cover numerous appearances of the target object.Incremental subspace learning algorithm uses a low-dimension subspace to represent the target and can deal with variations in pose,scale,and illumination.The proposed model uses a small number of PCA basis as the subspace templates and uses the trivial templates to model both reconstruction errors caused by sparse representation and the PCA basis representation.This strategy makes the model jointly exploit the advantages of both sparse representation and the incremental subspace learning.In addition,a customized APG method is proposed to solve the optimization problem effectively.Finally,a unified objective function is proposed to estimate the best position of the target.Experimental results show that the proposed method outperforms other state-of-the-art methods.2.A novel incremental subspace and probability mask constrained sparse representation model is proposed.The trivial templates in the sparse representation framework can be used to model the background pixels and noisy pixels.However,the trivial templates may be activated to model any pixels in any patch when there is no occlusion,which will lead to a decline of the recognition capability.Meanwhile,most sparse representation based tracking methods abstract the target feature from a bounding box.However,the shape of the target is usually not the same as the shape of the bounding box.This will result in including a set of unstable background pixels in the feature region,making the model harder to distinguish the target from background.To alleviate the above problems,the proposed model uses a probability mask to tell which pixels belong to target itself and which pixels belong to background and noise,then uses the probability mask to constrain the trivial templates.This probability mask makes the model penalizes more on those reliable pixels which likely belong to the target while penalizes less on those pixels which likely belong to background or occluded part of the target in order to improve the tracking accuracy.Experimental results have demonstrated that by using the probability mask,the proposed algorithm performs well in a wide range of tracking scenarios(e.g.,background clutter,motion blur,occlusion,scale variation).3.A robust object tracking algorithm based on sparse representation with convolutional features is proposed.Most of the traditional sparse representation based algorithms usually use the image intensity to construct the template set.However,the image intensity based trackers can hardly handle the complicated situation in practical visual tracking(e.g.,rotation,scale variation and pose variation)due to the lack of target's structural information.Convolutional features contains more information,which have stronger capability in describing the target.The low-level convolutional features contain more detailed inner information of the target which is useful to precisely locate the target,while the high-level convolutional features contain more sematic information which is insensitive to significant appearance change.To this end,the proposed model introduces convolutional features in sparse representation model and uses the convolutional features from different convolutional layers to improve the robustness.Experimental results show that this method has achieved favorable tracking performance comparing with traditional sparse representation based algorithms.