基于改进ORB-SLAM2算法的温室机器人定位与稠密建图方法

Greenhouse Robot Localization and Dense Map Building Method Based on Improved ORB-SLAM2

摘要:

针对温室内环境复杂，ORB-SLAM2算法无法构建稠密地图的问题，本文提出了基于改进ORB-SLAM2算法的温室机器人定位与稠密建图方法。首先，在跟踪线程中提出一种根据图像总体像素自适应调整特征点提取阈值方法，提升特征点提取的质量和数量；其次，在ORB-SLAM2算法基础上，结合帧间相对位姿计算，增加旋转量与平移量作为关键帧选择条件，降低关键帧数量和平均跟踪时间，提高定位精度；最后，引入稠密建图线程，通过点云恢复、统计滤波、点云拼接及体素滤波算法，融合多帧点云数据，生成精细三维稠密地图。为验证方法的有效性与实用性，分别进行公开数据集仿真分析与真实场景测试，在Freiburg1_room、Freiburg1_xyz、Freiburg1_desk序列上，改进ORB-SLAM2算法比ORB-SLAM2算法运行轨迹更接近真实轨迹，平均绝对轨迹误差分别降低46.00%、29.01%、39.85%。在3种不同枝叶遮挡的温室环境内，相比ORB-SLAM2算法，改进ORB-SLAM2算法特征点匹配数量分别平均提升7.20%、12.37%、12.81%；同时，关键帧平均数量分别从400、525、1132帧减少到371、411、708帧，且平均跟踪时间分别从0.0390、0.0357、0.0318s减少到0.0373、0.0343、0.0290s。试验结果表明，改进ORB-SLAM2算法的估计轨迹与温室机器人实际运动轨迹基本契合，具有良好的回环检测性能，准确还原了作物与过道在三维空间中的真实分布，成功构建出温室场景的三维稠密点云地图。该方法可为温室移动机器人的定位与导航提供技术支撑。

Abstract:

Aiming at the problem that the environment inside the greenhouse is complicated and ORB-SLAM2 cannot build dense maps, the research on the greenhouse robot localization and dense map building method was carried out based on the improved ORB-SLAM2. Firstly, a feature point extraction threshold method was proposed in the tracking thread to adaptively adjust the feature point extraction threshold according to the overall pixels of the image to improve the quality and quantity of feature point extraction. Secondly, on the basis of ORB-SLAM2, combined with the relative position calculation between frames, the amount of rotation and translation was added as the key frame selection condition, which reduced the number of key frames and the average tracking time, and improved the positioning accuracy. Finally, a dense map building thread was introduced to generate fine 3D dense maps by fusing multi-frame point cloud data through point cloud recovery, statistical filtering, point cloud stitching and voxel filtering algorithms. In order to verify the effectiveness and practicality of the method, simulation analysis of public datasets and real scenario tests were conducted, respectively, and the improved algorithm was closer to the real trajectory than the ORB-SLAM2 running trajectory on the Freiburg1_room, Freiburg1_xyz, and Freiburg1_desk sequences, and the average absolute trajectory error was reduced by 46.00%, 29.01%, 39.85%, respectively. Within the greenhouse environments with three different branch and leaf shading, the improved algorithm improved the number of feature point matched by an average of 7.20%, 12.37%, and 12.81% each compared with ORB-SLAM2;meanwhile, the average number of keyframes was reduced from 400, 525, and 1,132 frames to 371, 411, and 708 frames, respectively, and the average tracking time was reduced from 0.0390s, 0.0357s, 0.0318s to 0.0373, 0.0343, and 0.0290s, respectively. The experimental results showed that the estimated trajectory of the improved algorithm basically fit with the actual trajectory of the greenhouse robot, which had good loopback detection performance, and a three-dimensional dense point cloud map of the greenhouse scene was successfully constructed, which accurately restored the real distribution of the crops and aisles in the three-dimensional space. This method can provide technical support for the localization and navigation of greenhouse mobile robots.