Robustness Research Of Visual Tracking In Complex Scenes

Visual object tracking is one of the most important research topics in computer vision due to its crucial role in realistic applications,such as video surveillance,intelligent traffic monitoring system,vehicle navigation,pilotless driving,unmanned aerial.vehicle,human-computer interaction,intelligent robot system and more.Furthermore,video tracking technology is the fundament role of others computer vision research fields.For example,three-dimensional reconstruction,virtual reality and augmented reality all need the assistance of tracking technology.Visual tracking has recently gained significant progress,but object tracking is still a difficult problem,because a lot of factors influence the tracking accuracy and robustness,such as heavy occlusion,shape deformations,background clutter,rotation,scale variation,motion blur and low resolution,which make visual tracking complex and challenging.So it is of great significance to achieve robustness and long-term tracking in complex scenes.The main research works and results are as follows:(1)Aiming at the visual object loss problem in complex scene,a relocation method object tracking method is proposed.In this algorithm,a re-detection mechanism is adopted to relocate the loss target based on the correlation filter framework.First,a more stable target failure judgment criterion is designed to trigger the re-detection procedure.And then,support vector machine(SVM)classifier is exploited to get the refinement positive and adopt the online Passive-Aggressive learning algorithm to update it.Extensive experimental results on the OTB50 benchmark dataset show that the algorithm improve the object tracking accuracy in complex scene.(2)Aiming at the features is not rich and the robustness is not strong acquired by the traditional method,a correlation filter object tracking algorithm is proposed based on the deep learning Convolutional Neural Networks(CNNs)features and re-detection mechanism.The algorithm utilizes the better generalization ability of the deep learning features and uses the high-level semantic features and low-level hand-made Hog features to construct the appearance model features.In order to obtain long-term and robust tracking results,a re-detection mechanism is used to relocate the loss target.(3)Aiming at the inherent boundary effect and the assumption of the target response is fixed during the tracking process,a joint solution to the above problems is proposed.In addition,considering the influence of time consistency on tracking performance,a more robust appearance model is constructed to improve the robustness and accuracy of target tracking in complex scenes.ADMM algorithm is used to get a global optimal solution quickly.Experimental results demonstrate that the algorithm can better deal with the tracking issue in complex scene.(4)Aiming at the spatial regularization can reduce the target appearance information and the influence of the object context information on the tracking performance,an appearance model combining the context information of the current frame and the historical object appearance information was proposed.In order to acquire more target appearance,the algorithm not only considers the global context target information,but also considers the influence of the historical target information,that is,the temporal relevant information on the tracking performance.By building more comprehensive target appearance model to improve the tracking performance in complex scene.In addition,in order to obtain fast calculation and global optimal solution,ADMM algorithm is used.Experiments show that our algorithm can effectively improves the performance of the compared benchmark algorithm.The algorithm can improve the obj ect tracking accuracy in complex scene.(5)Aiming at the unstable tracking performance of the single scale tracking,a multi-scale sparse representation histogram object tracking algorithm is proposed.In order to solve the robustness tracking problem in complex scenes,a multi-scale sparse representation method is proposed to cope with the problem of unstable tracking performance and cannot contain enough target information.Furthermore,in order to decrease the model drift and tracking failure,an occlusion handling strategy is proposed.Based on the above improved measures,robust object tracking performance can be obtained.