Multi-Camera System Based Robot Localization

The localization of mobile robots is a fundamental problem in robotics research.At present,the localization method also plays an important role in the autonomous driving research.The mul-tiple fisheye camera system is usually equipped,which is particularly common in automatic park-ing vehicles in order to fully perceive the information of the surrounding environment.Therefore,the research of visual localization method based on multi-camera system is helpful to the popular-ization of autonomous vehicles.The single-sensor-based localization algorithm cannot overcome complex scene changes and cannot meet the vehicle's demand for robustness.In order to improve the applicability of the algorithm in actual scenarios,information from multiple sensors is usu-ally used for localization.How to use the image information of multi-camera and fuse it with the information of other sensors has become a mainstream research issue.This thesis is aimed at mobile robots equipped with multiple fisheye cameras,and integrates multi-sensor configurations of gyroscopes and wheel odometers.Based on the planar motion forms of mobile robots,corresponding visual localization algorithms are proposed to make them suitable for different scenarios and conditions.The main research results of this thesis include:1.Aiming at the map scene,a mapping and localization system based on a multi-camera system is proposed.After the multi-sensor configuration is used to collect multi-sensor data in indoor and outdoor scenes,monocular VO,multi-camera global optimization,loop-closure detection and optimization,and multi-sensor joint optimization are performed according to the coarse-to-fine process arrangement.Sensor calibration parameters also participate in non-linear optimization as optimization variables to obtain environment maps and more accurate sensor calibration parameters.The map is expressed in the form of sparse feature point cloud map.The localization system uses the environment map as a priori,performs a global search in the map according to the image descriptor,realizes global localization,and optimizes the localization results to obtain accurate and robust localization results.2.Aiming at the non-map scene,a visual odometry based on a multi-camera system is pro-posed,which is a robust and efficient motion estimation method suitable for vehicles in the condition of plane motion.According to the measurement model of wheel odometer and gyroscope proposed in this thesis,the pre-integration method of wheel odometer-gyro is derived,including discrete integration method and error transfer equation.We integrate the integration result with the reprojection error term obtained from the feature tracking of the camera image,and optimize it in the tightly coupled sliding window optimization frame-work,and estimate the camera pose,landmark points,and random walk bias of the sensor in real time.In addition,corresponding initialization methods and loop-closure detection methods are also proposed to maximize the advantages of existing sensors and improve the performance of the algorithm.