Abstract: Abstract:Light Detection And Ranging Simultaneous Localization and Mapping (LiDAR SLAM) technology is generally suitable for static environments. However, in dynamic scenes, the effectiveness of localization and mapping can be compromised. Ground segmentation modules are typically used for point cloud classification, but under-segmentation of the ground can affect the selection of feature points. Additionally, most frameworks rely on a single loop closure detection method, which can lead to missed detections. To address these issues, a LiDAR SLAM system that incorporates Ground Segmentation and Loop Closure optimization in dynamic environment was proposed(GSLC-SLAM).First, lmnet was employed to remove dynamic points in the point cloud. A distance image and a residual image were generated as inputs to the network, and the salsanext network was used to predict dynamic objects. Next, the efficient gridestimate algorithm was utilized for ground segmentation, where a non-uniform grid division method was applied to reduce the number of grids, thereby ensuring segmentation efficiency. Ground points were further filtered based on three criteria: orthogonality, height, and flatness. Finally, a new loop closure detection method was introduced, in which LinK3D(Linear Keypoints for Three Dimensions point cloud) descriptors were combined with a BoW3D(Bag of Words for Three Dimensions point cloud) bag-of-words model. Descriptors were generated from edge feature points, matched using a Hamming distance-like approach, and a BoW3D database was constructed to store LinK3D descriptors extracted from keyframes for loop closure detection.Experimental results on the KITTI (Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago Dataset) dataset demonstrate that compared to the Lego-Loam(Lightweight and ground-optimized Lidar odometry and mapping) algorithm, the proposed method reduces error by 5.8%, 78.2%, and 12.5% on KITTI sequences 00, 02, and 05, respectively. Compared to F-Lloam(Fast LiDAR odometry and mapping), the error reduction is 76.7% and 53.8% for sequences 00 and 05, although the results for sequence 02 are less favorable. After validation, the system presented can effectively reduce the interference of dynamic objects, accurately segment ground points, and minimize missed loop detections, resulting in higher positioning accuracy and robustness.

Key words: Keywords:dynamic detection, ground segmentation, loop detection, Light Detection and Ranging Simultaneous Location and Mapping (LiDAR SLAM), automatic driving

摘要： 摘 要：激光雷达同时定位与建图（LiDAR SLAM）技术通常适用于静态环境下，然而在动态场景下，定位与建图效果会受到影响。地面分割模块通常用作点云分类处理，但是地面欠分割问题会影响特征点的选择。通常的框架只使用一种回环检测方法，这可能会导致漏检现象。针对上述问题，提出动态场景下基于地面分割与回环优化的激光雷达定位与建图系统（GSLC-SLAM）。首先利用lmnet对点云进行动态剔除，该算法将生成的距离图像与残差图像作为网络的输入，通过salsanext网络预测出动态物体。其次，利用高效的gridestiamte算法进行地面分割，该算法利用了不均匀网格划分的方法，减少了网格的数量，从而保证了分割的效率，利用正交性、高度和平坦度三个指标进一步筛选地面点。最后使用LinK3D(Linear Keypoints for Three Dimensions point cloud)描述子与Bow3D(Bag of Words for Three Dimensions point cloud)词袋构成的新的回环检测方法检测回环，该算法利用边缘特征点生成描述子，使用类似于汉明距离的方式进行描述子匹配，并采用类似于词袋的方法构建BoW3D作为LinK3D描述子的数据库，对关键帧提取的描述子进行存储以及回环检测。在KITTI（Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago Dataset）数据集进行测试的实验结果表明，在KITTI00、02与05序列中与Lego-Loam(Lightweight and ground-optimized Lidar odometry and mapping)算法相比分别降低了5.8%，78.2%，12.5%，相比于F-Loam(Fast LiDAR odometry and mapping)，00与05序列分别降低了76.7%，53.8%，但是02序列结果不佳。经过验证，本文系统可以实现减少动态物体的干扰、精确分割地面点、减少回环检测漏检的目的，使系统定位精度更高更鲁棒。

关键词: 关键词：动态检测, 地面分割, 回环检测, 激光雷达同步时定位与建图(LiDAR SLAM), 自动驾驶