Odometry Estimation via CNN using Sparse LiDAR Data

3D depth sensors such as LiDAR have proven to be very useful in recognizing the surrounding environment for the past decade, but the methods using these 3D data directly is not very much. Especially, there is few methods exist to use deep learning because of the sparseness characteristic of the LiDAR data. This paper presents the odometry estimation of sparse LiDAR(3D laser scanning) data using deep learning. We first voxelize the given consecutive LiDAR scans and concatenate the voxelized data to make input tensor. Then we train the 3D convolutional neural networks with the tensor to calculate 6DoF pose between the LiDAR scans. The proposed method can replace position information of the wheel encoders of GPS data when the data is absenting.

I. Introduction

In these days, many studies using 3D LiDAR sensor are underway. 3D LiDAR sensor is a sensor that provides distance information more accurately than LiDAR. One of the most famous LiDAR sensors is Velodyne. The advantage of LiDARs with respect to cameras is that the noise associated with each distance measurement is independent of the distance and the lighting conditions. LiDAR-based studies include odometry estimation, point cloud registration, and so on. However, the LiDAR sensor has a disadvantage in that the information is sparse(see sample images in Fig. 1. You can see the points get more sparse as the distance increase), so matching is not easy. For this reason, most LiDAR approaches are variations of the traditional iterative closest point (ICP) scan matching which is a well-known scan-to-scan registration method [1], [2]. Many studies have also fused other sensors such as camera, IMU or GPS [3]–[5].