Fast And Robust Visual Tracking In Complicated Scenes

As one of the important research topics in computer vision,visual tracking has been widely used in many fields such as precision-guided munition,video surveillance,intelligent navigation and human-computer interaction.Although the research on the visual tracking algorithm has made great progress in recent decades,it is still challenging to achieve robust and fast model-free tracking due to the significant object appearance variation caused by deformation,illumination variation,and partial occlusion.In this paper,we focus on multiple feature integration,fast scale estimation and self-correction mechanism.A brief introduction of our paper is as follows:To improve the discriminability and robustness of the appearance model,this paper first analyzes the confidence of output response map,and proposes an adaptive weighted multiple feature fusion approach.By integrating two complementary features,the robustness and generalization ability of the appearance model have been significantly improved.To achieve fast and robust object scale estimation,this paper proposes a scale estimation method based on scale filter.The adaptive scale estimation method includes feature pyramid construction,scale filter training and scale feature dimension reduction.This method avoids the exhaustive searching in scale space,and effectively reduces the computational cost.Besides,the scale estimation module proposed in this paper is generic and can be incorporated into any tracking method without inherent scale estimation.To address the problem of error accumulation and tracking drifting,this paper first analyzes the boundary effects existed in correlation filter based trackers,and explains how the tracking drift occurs from a viewpoint of an open-loop system.Besides,this paper proposes an effective and generic self-correction mechanism by introducing a Peak-to-Sidelobe Ratio(PSR)and a step-wise searching approach.The proposed self-correction mechanism can effectively alleviate the problem of error accumulation and tracking drifting,and perform well when the object is under fast moving and occlusion.Finally,a high-speed object tracking system is developed based on a platform.The system is based on Android platform,and can utilize the captured video to achieve object tracking.Experimental results on publically available datasets and real scenes show that the system can achieve satisfactory performance and good robustness.