【作者】 李琦；

【导师】 秦兆博；

【作者基本信息】 湖南大学 ， 机械工程， 2023， 硕士

【摘要】 近年来,随着汽车行业的蓬勃发展以及国内人均汽车保有量的逐渐增多,泊车已成为广大司机面临的严峻问题。自主代客泊车技术(Autonomous Valet Parking,AVP)可以充分利用有限的停车资源,提高停车位利用率,缓解城市停车难问题。定位与建图是AVP系统不可或缺的核心模块。然而,针对地下停车场环境的建图定位尚存在许多亟待改善和解决的问题,主要包括停车场光线暗、特征少、纹理弱等。为了改善该环境下的定位性能,尤其是在无GPS信号的情况下,本文通过研究融合语义特征的视觉惯性SLAM算法,取得了良好的效果。本文的主要研究内容如下:(1)针对地下停车场环境的定位问题,提出了一种融合语义信息的视觉惯性定位算法框架。该算法首先通过视觉里程计和IMU预积分进行视觉惯性信息的融合,并完成视觉惯性初始化。同时,利用四个鱼眼摄像头输入图像构建全景环视图像,并采用语义分割算法提取停车场环境语义信息。然后根据视觉惯性紧耦合位姿完成逆投影变换,获得语义特征投影地图。最后采用回环检测和位姿图优化方式减小累积误差,完成全局位姿图优化,实现较高精度的定位效果。(2)针对多传感器融合数据处理问题,本文研究了视觉单目摄像头、多鱼眼摄像头以及IMU惯性传感器等多传感器之间数据融合处理过程。本文完成了视觉传感器和IMU传感器的标定及视觉-IMU联合标定,其中包括摄像头内参与畸变系数的标定矫正,IMU的内参标定等。此外,基于IMU测量模型,深入分析了IMU预积分的意义,并对IMU预积分过程做了详细的推导。此外,本文研究了鱼眼摄像头广角相机理论和联合标定过程,完成了鱼眼摄像头畸变矫正,逆投影变换,图像拼接融合等过程,最终构建了以车辆为中心的全景环视图像,为利用停车场全景环视下的语义信息提供了强有力的支持。(3)基于Gazebo仿真环境和地下停车场场景数据进行算法试验。为了验证本文所研究的面向停车场环境的视觉惯性SLAM定位算法的可行性,分别基于仿真环境和自动驾驶实验测试平台车进行实验测试。首先,在仿真条件下,本文采用仿真软件构建了地下停车场地面模型,测试该算法的可行性。此外,基于地下停车场真实环境对算法进行实验,并分析实验结果,以验证本文定位系统在地下停车场环境下的定位准确性和鲁棒性。更多还原

【Abstract】 In recent years,with the vigorous development of the automotive industry and the gradually increasing number of private cars owned by individuals in China,parking has become a serious problem for many drivers.Autonomous Valet Parking(AVP)technology presents a viable solution for optimizing the utilization of scarce parking resources,enhancing parking space efficacy,and mitigating the urban parking challenges.Localization and mapping are indispensable core modules of the AVP system.However,there are still many problems to be improved and solved in localization and mapping for underground parking environments,mainly including the dim light,few featu res,and weak textures of parking lots.In order to improve the localization performance in such an environment,especially in the absence of GPS signals,the paper has achieved good results by studying the visual-inertial SLAM algorithm fused with semantic features.The main research contents of this paper are as follows:(1)To address the localization problem of the AVP system in underground parking environments,a visual-inertial localization algorithm framework fused with semantic information is proposed.The algorithm first integrates visualinertial information through visual odometry and IMU preintegration,and completes visual-inertial initialization.Meanwhile,four fisheye cameras are used to input images to construct a panoramic surround image,and a semantic segmentation algorithm is used to extract semantic information of the parking environment.Then,inverse projection transformation is completed based on the tightly coupled visual-inertial pose,and a semantic feature projection map is obtained.Finally,loop detection and pose graph optimization methods are used to reduce cumulative errors,complete global pose graph optimization,and achieve high-precision localization.(2)The paper investigates the data fusion process among multiple sensors,namely a monocular camera,multiple fisheye cameras,and an IMU inertial sensor,to address the multi-sensor fusion data processing problem.The visual sensor and IMU sensor are calibrated separately and jointly,including the intrinsic and distortion correction calibration of the camera and the intrinsic calibration of the IMU.Furthermore,the paper analyze the significance of IMU preintegration based on the IMU measurement model and provide a d etailed derivation of the preintegration process.In addition,the paper study the fisheye camera theory and the joint calibration process,complete the distortion correction,inverse projection transformation,and image stitching fusion processes of fisheye cameras,and ultimately construct a panoramic surroundview image centered on the vehicle,providing robust support for utilizing semantic information in the panoramic surround-view of parking lots.(3)Algorithm testing was conducted based on Gazebo simulation environment and underground parking lot scene data.In order to verify the feasibility of the visual-inertial SLAM localization algorithm for parking lot environments studied in this paper,experiments were conducted on both simulation environments and a self-driving experimental test platform vehicle.Firstly,a ground model of the underground parking lot was constructed using a software simulation environment to test the feasibility of the algorithm under simulation conditions.Then,using real-world data from an underground parking lot,the algorithm was tested and the experimental results were analyzed to verify the accuracy and robustness of the localization system in underground parking lot environments.更多还原