Design And Implementation Of Semi-direct Based Monocular Visual Odometry

With the continuous evolution of computer vision and image processing technology in recent years,visual sensors have been used in more and more scenes,and odometry based on visual sensors have also received increasing attention from researchers.Visual sensors are becoming one of the main sensors in odometry applications due to their low cost and rich information.Compared with visual based Simultaneous Localization and Mapping,visual odometry has better efficiency and lower hardware requirements.Wheels may slip and idling on rough roads,and traditional odometer based on wheels has large error,while visual sensor based odometry do not influence by this.The goal of this paper is to design and implement a robust,fast,and efficient visual odometry system that is suitable for platforms with limited computing resources.Combining the theoretical basis of direct method and feature based method,consider both robustness and real-time,this paper proposes a semi-direct monocular visual odometry that combines line features and point features.The main work of this article can be summarized as:(1)The development and research status of visual odometry technology at home and abroad are summarized.The research significance of this paper is discussed.The relevant theoretical knowledge of monocular visual odometry is deeply studied,which lays a foundation for the system design and algorithm improvement of the subject.At the same time,the principle,advantages and disadvantages of the feature based point method,direct method,and semi-direct method are analyzed,which lays a foundation for the system design and algorithm improvement of the subject.(2)Based on the design idea of semi-direct method,this paper propose a semidirect monocular visual odometer system combining point features and line features which improve and extended the SVO(Semi-direct Visual Odometry).The system includes four modules: image preprocessing,model selection based initialization,semidirect method based camera tracking,depth filtering based mapping.The proposed system captures the advantages of SVO's fast and efficient,and at the same time it improves the defects of SVO's lack of robustness,and can make a quick,robust,and high-precision estimation of the camera's motion.(3)This paper proposes targeted improvement measures for SVO's initialization,pose estimation,map construction,and other defects,including six aspects: Introduce line fearture to improve the robustness of the scene with few point feature;Use model selection to improve success rate and initial pose quality of initialization;A new key frame selection strategy is proposed to reduce the restrictions on the camera's viewing angle,thereby reducing the restrictions on the application scene;A pre-update strategy is proposed to improve the convergence speed of depth of 3D points;The method of updating the map point reference frame is proposed to reduce the sensitivity to ambient light changes.Finally,the experimental verification of the system is carried out on the public data set.The experimental results show that the system designed in this paper has excellent performance in terms of real-time performance,robustness,and precision of camera trajectory.