【作者】 马岩；

【导师】 张亮；

【作者基本信息】 山东大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 移动机器人是人工智能技术落地应用的重要载体,而定位和建图是移动机器人执行智能化任务的前提。由于相机价格低廉且可以获取到丰富的场景信息,因此基于视觉的同步定位与建图(Simultaneous Location and Mapping,SLAM)算法得到了广泛的应用。但在光照不佳、快速运动等无法有效提取和正确匹配特征点的挑战性场景中,如何确保定位与建图的精度与鲁棒性,是目前该领域的关键瓶颈问题。因此,本文设计并实现一个多传感器融合SLAM系统,使用双目相机结合惯性测量单元(Inertial Measuring Unit,IMU)搭建视觉惯性融合里程计进行定位,并结合使用激光雷达进行同步建图,实现在光照变化、快速运动等挑战性场景下高精、鲁棒的同步定位与建图。主要研究工作包括以下几个部分:为保证基于特征点的多传感器融合SLAM系统前端特征提取和匹配的精度和鲁棒性,对ORB(Oriented FAST and Rotated BRIEF,ORB)特征点的提取方法进行研究。首先,为解决ORB特征点对光照敏感的问题,使用基于自适应的Gamma校正和限制对比度的自适应直方图均衡法的图像预处理方法,平衡图像亮度的同时提高图像对比度;此外,提出基于谷值估计和双阶段多阈值大津法的ORB提取算法,通过引入图像分割算法对图像进行处理,挖掘潜藏在图像内的细节,并使特征点之间的区别更加明显,从而提高特征点检测和匹配的准确性。实验结果表明,所提出的方法在Leuven数据集中提取特征点数比ORB算法多了 1.5倍,匹配精度最高提高了 18.82%。为提高多传感器融合SLAM系统前端特征跟踪的鲁棒性,提出融合IMU信息的LK光流法。该方法使用基于图像金字塔的光流法跟踪特征点,并通过引入IMU信息辅助预测追踪特征点的初始值,在EuRoC数据集中点特征的追踪长度可达22.8帧,相对于Lucas-Kanade 光流法提高了 6.4帧;为了解决基于特征的SLAM系统前端追踪线程中描述子计算耗时较高而利用率较低的问题,采用在普通帧中不进行描述子计算,只在关键帧中提取描述子的方案,以此提高前端的计算效率。为增强系统对于传感器异常的鲁棒性,提出基于动态权重的多传感器融合后端方案。首先,详细推导IMU和相机测量值的残差和雅可比矩阵,并利用图像的内点率动态调整相机在优化问题中的权重;此外,为了消除累积误差,使用基于词袋模型的回环检测为优化问题提供回环约束;最后,使用激光雷达建立可以被机器人用来导航和避障的栅格地图,实现多传感器融合的同步定位和建图。在EuRoC数据集和TUM-Ⅵ数据集的20个序列中进行对比实验,结果表明,在光照变化、快速运动、部分视觉特征缺失的挑战性场景中,相较于基于点特征或光流法的相应算法,本文算法均实现了更精确的定位。在实际场景中,通过模拟光照变化,使用移动机器人验证了本文设计系统在光照变化场景中的可以实现鲁棒的同步定位与建图。更多还原

【Abstract】 Mobile robot is an important carrier for the application of artificial intelligence technology,while localization and mapping are the prerequisites for mobile robots to perform intelligent tasks.Since cameras are inexpensive and can acquire rich scene information,vision-based simultaneous location and mapping(SLAM)algorithms have been widely used.However,how to ensure the accuracy and robustness of localization and mapping in challenging scenarios,such as poor lighting and fast motion,where feature points cannot be effectively extracted and correctly matched,is the key bottleneck in this field.Therefore,in this thesis,we design and implement a multi-sensor fusion SLAM system using binocular camera combined with inertial measuring unit(IMU)to build a visual inertial fusion odometry for localization,and combined with LIDAR for simultaneous mapping,to achieve high-precision and robust simultaneous localization and mapping in challenging scenarios such as poor illumination and fast motion.The main research work includes the following parts:To ensure the accuracy and robustness of the front-end feature extraction and matching of the feature-point-based multi-sensor fusion SLAM system,the ORB(Oriented FAST and Rotated BRIEF)feature point extraction method is investigated.First,to solve the problem of ORB feature points’ sensitivity to illumination,an image preprocessing method based on adaptive Gamma correction and adaptive histogram equalization method for limiting contrast is proposed,which balances the brightness of the image and improves the contrast at the same time;in addition,an ORB extraction algorithm based on the valley estimation and two-stage multi-threshold Otsu method is proposed,and an image segmentation algorithm is introduced to process the image to excavate the details hidden within the image and make the difference between feature points more obvious,thus improving the accuracy of feature point detection and matching.Experimental results show that the proposed method extracts 1.5 times more feature points than the ORB algorithm in the Leuven dataset,and the matching accuracy is improved by up to 18.82%.To improve the robustness of the front-end feature tracking of the multi-sensor fusion SLAM system,the LK optical flow method that fuses the IMU information is proposed.The method uses the image pyramid-based optical flow method to track feature points,and aids in predicting the initial values of tracked feature points by IMU information,and the tracking length of point features in the EuRoC dataset can be up to 22.8 frames,which is an improvement of 6.4 frames with respect to the Lucas-Kanade optical flow method;to address the problem of high time-consuming but low utilization of descriptor computation in the front-end tracking thread of feature-based SLAM systems,the scheme of not computing descriptors in ordinary frames but extracting descriptors only for key frames is adopted to improve the computational efficiency of the system.To enhance the robustness of the system to sensor anomalies,a multi-sensor fusion back-end scheme based on dynamic weights is proposed.First,the residual and Jacobi matrices of the IMU and camera measurements are derived in detail,and the camera weights in the optimization problem are dynamically adjusted using the in-point rate of the image;furthermore,to eliminate the cumulative error,the bag-of-words model-based loop closure detection is used to provide loop constraints for the optimization problem;finally,the LIDAR is used to build a raster map that can be used by the robot for navigation and obstacle avoidance,and to achieve the synchronization of the multi-sensor fusion localization and mapping.Comparative experiments in 20 sequences of the EuRoC dataset and the TUM-VI dataset show that the algorithms in this thesis achieve more accurate localization compared to the corresponding algorithms based on point features or optical flow in challenging scenarios with varying illumination,fast motion,and partial absence of visual features.In real scenarios,the system designed in this thesis can achieve robust simultaneous localization and mapping in light changing scenarios by simulating light changes and using a mobile robot.更多还原