Research On Visual Object Tracking Of Mobile Robot In Complex Environment

Visual object tracking has received increasing attention over the last decades and has remained a very active research direction.It has a large range of applications in diverse fields like surveillance,human-computer interactions,robotic and automatic driving.With the real-time and stable target tracking function,the mobile robot can realize humancomputer interaction and help people complete complex and heavy tasks,such as tracking person to complete the handling of luggage or goods.However,the tracking algorithm is challenged by the complex conditions such as occlusion,out-of-view,and illumination variation.Therefore,this thesis investigates the problem of model corruption,full occlusion,out-of-view in a discriminative correlation filter(DCF),and propose a novel tracking algorithm.Finally,the novel tracking algorithm is transplanted to the mobile robot platform and tested in a real and complex environment.The experimental results demonstrate that the proposed tracker algorithm has good accuracy for full occlusion and out-of-view.To summarize,the main contributions of this work are listed below in threefold:(1)A tracking uncertainty estimation method combining response confidence and peak-to-side ratio is proposed.Most of the DCF’s trackers ignore the reliability of the tracked results and lack an effective mechanism to refine the unreliable results.So,to overcome these issues this thesis proposes tracking uncertainty estimation method composed of peak-to-sidelobe ratio(PSR)and tracking confidence.At the same time,an adaptive model updating method is presented to solve the issues of model corruption and over-fitting in the traditional DCF trackers.(2)A re-detection method based on fully-correlational is proposed to avoid the target loss problem because of the full occlusion and out-of-view.DCF trackers still struggle in full occlusion and out-of-view scenarios due to the absence of a re-detection component.Therefore,in order to cope these issues this thesis proposes a fully-correlational filter that is able to re-detect the target in the whole image efficiently.The proposed fullycorrelational re-detection module can be integrated into DCF trackers to consistently boost the performance.This thesis performs comprehensive experiments on the challenging OTB100 and UAV20L datasets.The experimental results demonstrate that the proposed algorithm can meet the real-time requirements of mobile robot platform only using CPU,and has good tracking accuracy compared with the state-of-the-art trackers.(3)Based on the Turtlebot3 platform,a visual object tracking system for mobile robot is built,and the proposed tracking algorithm is verified by experiments in real scenes.The experimental results show that the proposed tracking algorithm performs robustly and accurately in the real scenes of the complex environment,such as full occlusions and out-of-view.