GR-LOAM: LiDAR-based sensor fusion SLAM for ground robots on complex terrain

Highlights

• Odometer increment model fuses IMU and encoder data to estimate robot pose on manifold.

• Abnormal sensor data detection and optimization weight adjustment.

• A multi-sensor ground-constrained refinement method with dynamic objects being removed.

• Extensive experiments based on a real ground robot in indoor and outdoor environments.

Abstract

Simultaneous localization and mapping is a fundamental process in robot navigation. We focus on LiDAR to complete this process in ground robots traveling on complex terrain by proposing GR-LOAM, a method to estimate robot ego-motion by fusing LiDAR, inertial measurement unit (IMU), and encoder measurements in a tightly coupled scheme. First, we derive a odometer increment model that fuses the IMU and encoder measurements to estimate the robot pose variation on a manifold. Then, we apply point cloud segmentation and feature extraction to obtain distinctive edge and planar features. Moreover, we propose an evaluation algorithm for the sensor measurements to detect abnormal data and reduce their corresponding weight during optimization. By jointly optimizing the cost derived from the LiDAR, IMU, and encoder measurements in a local window, we obtain low-drift odometry even on complex terrain. We use the estimated relative pose in the local window to reevaluate the matching distance across features and remove dynamic objects and outliers, thus refining the features before being fed to a mapping thread and increasing the mapping efficiency. In the back end, GR-LOAM uses the refined point cloud and tightly couples the IMU and encoder measurements with ground constraints to further refine the estimated pose by aligning the features on a global map. Results from extensive experiments performed in indoor and outdoor environments using real ground robot demonstrate the high accuracy and robustness of the proposed GR-LOAM for state estimation of ground robots.

Introduction

Simultaneous localization and mapping (SLAM) is essential for the navigation of ground robots in unknown environments. SLAM allows to estimate the robot pose and provide information for path planning. As cameras mounted on ground robots are usually small, affordable, and easy to install, they are widely adopted for SLAM. In recent years, many studies have achieved high-performance visual SLAM [1], [2], [3], and by fusing measurements from inertial measurement units (IMUs) and cameras, the positioning accuracy and robustness have been further improved [4], [5], [6]. Although visual SLAM has many advantages, camera images are sensitive to lighting, textures, and large variations in the viewing angle. As an active sensing method, LiDAR provides high-quality distance information as point clouds of their surroundings and is insensitive to light, being able to operate even at night. LiDAR odometry and mapping (LOAM) [7], [8] has achieved high performance and remained at the forefront of the KITTI vision benchmark ranking. From this development, many high-performance LiDAR SLAM algorithms have been derived [9], [10], [11].

Although the accuracy and robustness of LiDAR SLAM are higher than those of visual SLAM, some problems remain to be addressed. First, the movement of a ground robot distorts the LiDAR point cloud, leading to errors in point cloud matching. Second, LiDAR can only obtain few features in open environments and captures dynamic objects (e.g., vehicles, pedestrians) for mapping. Third, in environments such as long corridors, most points are located in similar planes, hindering localization and mapping. Fourth, given the slow update rate of LiDAR measurements (around 10 Hz), they cannot satisfy the feedback frequency required by pose controllers during robot navigation (usually above 50 Hz).

As individual sensors tend to flaw in real settings, ground robots are often equipped with multiple sensors (e.g., LiDAR, camera, IMU, encoder), and combining their information may improve the overall system performance. For instance, the IMU can measure angular velocity and linear acceleration, being more robust under dynamic motion and able to assist LiDAR data for point cloud distortion removal. However, the noise and random walk of IMU measurements increase the error over long integration periods. Although an encoder can measure the rotation angle of the robot wheel and allow to estimate the robot displacement, the estimation is only accurate on flat ground. Therefore, the suitability of encoders is limited to applications such as service robots and industrial automatic guided vehicles that are deployed in environments with flat ground. For other applications such as rescue, security, and inspection, robots often face complex terrain that includes slopes, steps, and stairs, as shown in Fig. 1. In such scenarios, sensor fusion comprising LiDAR, IMU, and encoder measurements is challenging. Moreover, given the inherent error of each type of sensor, inappropriate fusion may render the fusion worse than using the sensors separately. Therefore, proper sensor fusion should be devised to improve the overall localization and mapping performance.

We propose GR-LOAM, a method to fuse LiDAR, IMU, and encoder data for accurate and robust SLAM of ground robots on complex terrain. We propose an odometer increment model that allows fusing IMU and encoder data to calculate the robot pose variation on a manifold. In a local sliding window, odometry is then estimated by tightly coupling the LiDAR, IMU, and encoder data, and low-drift results are obtained even in challenging environments. In addition, we refine features by using the odometry results to remove dynamic objects and outliers, ensuring high feature quality before mapping as well as increasing the matching efficiency. Finally, GR-LOAM uses the refined feature points, tightly coupling IMU and encoder data again and including ground constraints to optimize the global robot state by aligning the features with a global map. We can summarize the contributions of the proposed GR-LOAM as follows:

• An odometer increment model fuses IMU and encoder data to estimate the pose variation on a manifold.

• Odometry based on factor graphs to tightly couple LiDAR, IMU, and encoder data allows to detect abnormal data and adjust the optimization weight of each sensor.

• A multi-sensor ground-constrained refinement method further optimizes the pose estimation and generates a global map with dynamic objects being removed.

• Extensive experiments based on a real ground robot in indoor and outdoor environments demonstrate the accuracy and robustness of the proposed GR-LOAM for localization and mapping of ground robots.

The remainder of this paper is organized as follows. Section 2 presents a survey of related works. Sections 3 Preliminaries, 4 Proposed GR-LOAM detail the proposed GR-LOAM. The experimental setup and results are reported and analyzed in Section 5. Finally, we draw conclusions in Section 6.