Research And Implementation Of Target Following Robotic System Based On Monocular Vision

Robotic object following system which is a systematic technology that relates to motion control of robot,computer vision,sensor fusion,machine learning,etc,has great research value and a wide range of application prospect in robot navigation,human-computer interaction,unmanned driving and disaster assistance.Vision-based object following robots can adapt complex and volatile environment,and they have broad,long visual field.These robots have higher accuracy on object recognizing than common rader-based sensors.This paper summarizes both advantage and weakness of target following system which based on laser and other sensors.Based on analysis of camera model,deep learning-based scene segmentation and object tracking algorithm,A*pathfinding algorithm,this paper introduces the monocular vision based target following system in detail,including how to achieve object tracking and scene recognizing,how to improve and optimize the pathfinding category based on A*algorithm to promote the efficiency of robot and the ability to process complex environment.The main contents and contribution of this paper are as follows:Firstly,on the basis of further understanding and comparision between existing algorithms and models,the GOTURN-based algorithm and SegNet-based model are implemented and optimized for object tracking and image segmentation.Extra datasets are captured and annotated,and the model is trained again.The modules of object tracking and scene segmentation can run in real time.Secondly,aiming at the problem of lacking spatial coordinates and unable to build grids map,this paper proposes to represent the 3D spatial position of target with its projection on the plane.All the spatial points are projected to the same plane.The homography of camera is calculated with calibration algorithm and iterative optimization.The pixel coordinates and world coordinates are associated with the homography.Thirdly,with combination of improved and optimized object tracking algorithm and target localization algorithm,long-range target following is achieved.To promote the efficiency and reduce the complexity of the system,this paper proposed to construct the grids map in polar coordinates system.The output of scene segmentation module is also projected to the grids map.With extensive experiments indoors and in the hallway,we demonstrate the real-time and efficiency of the proposed robotic system.