Session D1: Alternative Technologies for GNSS-Denied Environments

Radar-Inertial Navigation for Unmanned Ground Vehicles in GNSS-Denied Off-Road Environments
Petar Mitrev and Mohamed Atia, Carleton University
Date/Time: Wednesday, Sep. 18, 8:57 a.m.

Peer Reviewed

Accurate localization is critical to the successful navigation of unmanned vehicles. When GNSS signals are unreliable, such as in cluttered forests, indoor spaces, underground mines, or in a hostile situation where signals can be jammed and spoofed, localizing with other a combination of proprioceptive and exteroceptive sensors have shown to be effective. However, these localization algorithms heavily rely on sensors such as camera and LiDAR, which are significantly degraded during periods of harsh weather, poor lighting conditions, etc. Additionally, LiDAR is easily detectable if a vehicle is operating in hostile environments. Due to these limitations, using radar as the primary sensor for localization has recently received increasing attention. Radar, however, exhibits its own set of challenges, namely highly sparse and noisy point clouds compared to LiDAR. Existing radar odometry techniques include the combination of gyroscopes, IMUs, encoders, as well as automotive and scanning radars doppler velocities and point clouds. These techniques are typically restricted to indoor or structured city environments. In contrast, our research features an unmanned off-road ground vehicle operating in forested and off-road environments. We design and test a localization system capable of operating in these environments using automotive radar, IMU, and wheel odometry. We use both radar velocity measurements as well as a point cloud data via a scan densification and noise removal pipeline when sufficient environmental features are available, for example in an underground mine. When these features are not present, such as forests with uniformly dispersed trees, we operate without point cloud measurements and accurately navigate using only radar ego-velocity and our auxiliary sensors. Our objective is therefore summarized as the design and implementation of a navigation system for an unmanned ground vehicle in off-road environments such as dense forests or underground mines where GNSS signals are unavailable. Additionally, circumstances which require an unmanned vehicle to remain undetected by potentially hostile parties likewise prohibit the use of LiDAR. To overcome these problems, we use a combination of Frequency Modulated Continuous Wave (FMCW) radar sensors, IMU and driveline data to accurately localize an unmanned ground vehicle in restricted environments.
Our localization system is based on an EKF with an iterative update stage and adaptive measurement noise via covariance matching (IAEKF). We apply a weighted-least-squares velocity regression to the doppler velocity of radar targets where points are weighted according to their azimuth from the radar's midpoint. The regressed body velocity of the radars are then fused with IMU data. Additionally, we fuse forward speed from encoders as well as Zero Update Velocity constraints during stationary periods. Where possible, we also use the radar point cloud measurements and implement a scan matching procedure. To do this, we utilize the results of the doppler velocity regression previously performed to identify and remove outlier points from the individual radar scans. To further remove speckle noise and multi-path reflections, we then use the radar velocity and gyroscope to estimate a rough short-term odometry of the vehicle to overlap successive scans and cluster true environmental features. A density-based clustering algorithm is used to remove noise and leave the real environment points in the point cloud. We then perform a point-to-plane ICP in a scan-to-multi-scan fashion to compute changes in displacement and orientation. Data was collected in a densely forested area, an underground mine, and in a local outdoor testing ground. The platform consists of an 8-wheeled differential drive vehicle with an onboard computer for data collection. The vehicle is outfitted with an Xsens IMU, two automotive grade FMCW radars (one in each of the front corners), a Novatel GPS and Velodyne LiDAR for ground truth, and wheel encoders. The extrinsic (distance and orientation) between the IMU and radars were manually measured and further refined offline using the recorded data. Algorithm design and testing was done offline using MATLAB. It will be implemented in C++ for real-time use in the future.
In the environments with reliable features, i.e. underground mine and outdoor testing facility, the multi-scan density-based clustering produces clean and dense scans, alleviating the challenges radar faces compared to LiDAR. We observe a 3.4m RMSE error for the radar-IMU EKF and a 0.9 RMSE error for our augmented IAEKF with point-to-plane ICP. We see improvements in regions of hard cornering where the radar-based ICP aids in in translation estimates compared to integrated velocities. In densely forested areas, there is a severe lack of persisting features in the point clouds, which renders any form of scan matching not feasible. However, due to the high density of stationary point targets (trees) with similar returned power levels, a highly reliable radar ego-velocity can be obtained as this measurement is computed by the doppler readings in a single radar scan rather than tracking points across multiple scans. In these situations, the state estimate relies on the filter using just radar velocity, IMU, and wheel odometry data. To augment the performance of the filter in these environments, we implement the adaptive measurement covariance estimates as previously described. We observe a 3.3% RMSE for the radar-IMU EKF and a 1.8% RMSE for the IAEKF.
In conclusion, we observe the performance and feasibility of using automotive radar for the localization of off-road vehicles in GNSS denied regions. To our knowledge, this is the first application of FMCW radar to the navigation of an off-road ground vehicle in forested and underground environments via IAEKF, as well as the fusion of vehicle ego-velocity, point-to-plane radar-based ICP, and wheel odometry. We observe improvements in localization accuracy compared to previous techniques that only use IMU and radar velocity in an EKF formulation without point cloud data. We also observe improvements in performance to prior IMU and radar velocity approaches by formulating our EKF to be iterative as well as adaptive in environments which can only use radar velocity data. Our results show the accuracy and feasibility of using FMCW radar as the main localization sensor for off-road vehicles in GNSS denied, or hostile environments. Future work will focus on the implementation of our techniques to even longer spans of GNSS denied navigation.