Research On Robust Tracking Algorithms With Deep Visual Semantic Information

Visual tracking is a significant study in the field of computer vision.It has been widely applied in the domains of defense military,industrial production and social public security.However,visual tracking is challenging because it has been suffering from the influence of various factors,such as illumination,occlusion and deformation.Hence,visual tracking is a difficult topic in the field of computer vision.We systematically analyze the existing visual object tracking methods,among which sparse representation based tracking algorithms show better tracking performance when the number of candidates is limited.However,with the increasing of candidates number,the speed of solving sparse coefficients becomes very slow,which leads to heavy computational cost and decrease the tracking speed.Recently,deep learning has a huge impact on the field of computer vision.Deep models automatically learn object-related visual features from large scale data.We combine deep features with sparse representation model to select candidates that most related to the tracking object.While reducing the computational cost,the tracking precision and model robustness is improved.The innovative achievements are summarized as follows.1)We propose a robust object tracking algorithm that combines deep visual features with local sparse representation model.Deep features are sparse and selective.In the process of deep feature extraction,the neurons of convolutional layer are selective in response to the targets.According to the response strength of target and background region on the feature map,we can pick the neurons who are more interest in targets.After selecting neurons,object-related samples have higher response on feature maps than other samples.Thus,we can pick a small amount of object-related candidate samples.The selected ones will be verified by sparse representation model to determine the tracking object.The proposed method greatly reduces the number of candidates and improves the detection speed and precision by selecting with deep features.Experiment results show that the proposed method is effective in improving tracking speed and accuracy.2)We propose a robust object tracking algorithm,which combines deep visual semantic features.Based on previous study on exploiting single layer deep feature,we add semantic information,which combines deep features of one high and one low layer with sparse representation model for tracking.The low layer features of deep model carry more object-specific discriminative information which separates object and background.While,semantic features of high layer contain category information,which is also robust to partial occlusion.The proposed method handles the response amplitude difference of two layers of the same target with a balance parameter.Afterwards,we pick a small amount of object-related candidate samples and exploit sparse representation model to verify the selected samples.Experimental results show that the proposed method further improved the tracking performance,and has strong robustness and stability.