The Study Of Semantic-Supervised Templates For Robust Visual Object Tracking

As one of the most fundamental tasks in computer vision,visual object tracking is of great significance for intelligent analysis of video images,and it has broad prospects and needs in many applications,such as video surveillance,unmanned driving and human-computer interaction.Visual object tracking is a process of accurately locating a specified object in the image sequence or video,and it is a very challenging task while it faces various difficulties such as illumination change,occlusion,deformation and background clutters during tracking.In this thesis,a unified model for visual object tracking task is proposed based on saliency detection and similarity measure.Under the model,two enhanced target templates are designed to learn more robust target representations,to solve various difficulties during tracking,and finally achieve more accurate tracking.The main contributions in thesis are as follows:1.A unified mathematical model for visual object tracking task is proposed based on saliency detection and similarity measure.Inspiring by the siamese network two enhanced target templates are designed to achieve more robust target representations.One is the context-aware template,which is to learn the salient features from more contextual information while considering the foreground and the background as a whole,and the second is the semantic-supervised template,which is to obtain an enhanced target-only template by semantic supervision learning,it achieves a more robust target representation after combining the whole template of the foreground and the background efficiently.Both of the two trackers,which are proposed based on the above enhanced templates,obtain more than 1.5% performance improvement on OTB dataset preliminarily.2.To improve the performance of the semantic-supervised template based tracker,self-adaptive fusion mechanism of template matching responses is proposed to merge the response of the enhanced target-only template and the whole template of the foreground and the background efficiently,at the same time,multi-layer fusion feature is introduced and optimized to take full advantage of the shallow appearance feature and the deeper semantic feature within deep networks.The improved tracker obtains more than 5%performance improvement on OTB dataset,and achieves real-time tracking on a single graphic processor.3.Based on the above improved tracker,a high-performance tracker is proposed to preliminarily explore the problem of efficient tracking in the case of inaccurate initialization of the target which is to be tracked,and it achieves robust tracking under inaccurate target initialization.