HRegNet-LO:基于端到端深度神经网络的LiDAR里程计测量

HRegNet-LO: LiDAR Odometry Measurement Based on End-to-End Deep Neural Network

针对传统LiDAR里程计(LO)测量方法在处理无初值、长序列点云配准时存在精度低、稳定性差等问题,本文引入端到端点云配准网络(HRegNet),提出一种基于深度神经网络的LO测量方法——HRegNet-LO算法,以期实现更加准确、鲁棒的LO测量。所提算法由两个核心模块组成:前端计算和后端优化。在前端scan-to-scan配准中,主要是依据原始点云的3D坐标,采用HRegNet网络,计算出相邻两帧点云的初始转换矩阵,实现LO初始位姿计算;在后端scan-tomap配准中,主要是通过提取特征点构建特征地图,应用迭代最近邻点(ICP)算法,每间隔一定距离对初始位姿进行优化,以减小预估轨迹中的漂移。在Kitti odometry数据集上对所提算法的性能进行了评估,并与LOAM、F-LOAM等算法作对比分析。实验结果表明,所提算法相对旋转、平移误差分别在0.003°/m和1%左右,每帧位姿计算耗时约为100 ms,可以满足LO测量对于精度和实时性的要求。

Objective Vehicle odometer plays an important role in unmanned driving systems,and its main task is to determine the position and pose of the vehicle.The calculation method falls into two categories of visual odometry(VO)and LiDAR odometry(LO).VO is limited by illumination conditions,while LO is unaffected by illumination changes through active laser beam emission,making LO more suitable for vehicle odometry measurement than VO.However,the traditional LO algorithms face the problems of low accuracy and weak robustness when they are employed to register the continuous LiDAR frame point clouds without initial value and long sequence.We introduce an end-to-end point cloud registration network HRegNet and propose an LO computation algorithm using deep neural networks HRegNet-LO. Such a LO computation strategy exhibits strong robustness, good accuracy, and high efficiency, and we hope that the proposed algorithm can be helpful to the LO measurement.Methods The HRegNet-LO algorithm includes two core modules of front-end calculation and back-end optimization.The calculation is shown in Fig. 1.In frond-end calculation, scan-to-scan registration is conducted to obtain the transformation of adjacent two LiDAR frame point clouds. Firstly, HRegNet is employed to calculate a rough transformation matrix between two adjacent LiDAR frames without initial value by adopting the original point clouds. Secondly, according to point curvature, some feature points belonging to different categories are extracted from the original point clouds. Thirdly, the rough transformation matrix is refined by an iterative closest point (ICP) algorithm using feature points. Fourthly, LO initial pose is obtained by the sequence registration of the transformation matrix between two adjacent LiDAR frames.In back-end optimization, scan-to-map registration is conducted to adjust the LO initial pose. Firstly, a feature map is constructed using the initial LO pose and feature points. Secondly, by finding the corresponding feature points from the current LiDAR frame and feature map, a pose correctional value is obtained via the ICP algorithm. Thirdly, the LO pose of the current LiDAR frame is adjusted to reduce the estimated trajectory drift.Results and Discussions Comprehensive experiments are carried out on the Kitti odometry dataset to evaluate and demonstrate the performance of the proposed HRegNet-LO algorithm. The LOAM, F-LOAM, and some other methods are compared with our method.Figures 5 and 6 show the qualitative results of HRegNet-LO on three experimental datasets. Figure 5 indicates that the measurement results of the HRegNet-LO algorithm along the x and z axes are in good agreement with the real pose information of the ground truth, while the measurement accuracy along the y axis is poor. This is because the point clouds obtained by the vehicle-mounted LiDAR system have more constraints in the horizontal direction (along the x and z axes),but fewer constraints in the vertical direction (along the y axis). Therefore, the positioning accuracy of the horizontal direction is better than that of the vertical direction. In practical applications, the horizontal component of LO measurement results is more important than the vertical component. To observe the horizontal measurement results of the proposed algorithm more clearly, Fig. 6 shows the aerial view of our algorithm's measurement results on the three testing datasets, where we can find that the horizontal component of the measured results is very close to the real path of the ground-truth. To conclude, the proposed algorithm can accurately realize the LO measurement only by LiDAR point clouds.Quantitative and comparison experiments are also conducted, whose results are shown in Fig. 7 and Table 2. In terms of relative rotation error (RRE) and relative translation error (RTE), our method almost obtains the best LO measurement results, and the RRE and RTE per 100 m can be controlled at about 0. 3° and 1 m respectively, which can meet the positioning accuracy requirements in unmanned driving.Ablation experimental results are shown in Fig. 8 and Table 3, which show that the back-end optimization can improve the accuracy by about 70%.Conclusions We design a novel LO measurement method HRegNet-LO algorithm. The front-end calculation is realized by HRegNet deep neural network, and back-end optimization is realized by point clouds feature map. Experimental results show that the RRE and RTE of HRegNet-LO are about 0. 003°/m and 1% respectively, and the consumed time of calculating each frame pose is about 100 ms. The proposed method can satisfy the accuracy and real-time requirements in LO measurement. However, we only consider using 3D coordinates of LiDAR point clouds to achieve LO measurement.Multi-system fusion methods such as integrated inertial navigation, vision, and LiDAR odometer measurement technology can be considered in subsequent studies to improve the reliability and accuracy of LO measurement results. Meanwhile,the generalization of deep neural networks also should be improved.