This package generalizes ekf_rio and ekf_yrio for multi radar sensor setups and provides a faster implementation using approximated radar clones. We used an instantaneous approach which can provide yaw aiding without scan matching. An evaluation with the Multi Radar Inertial Datasets JGN 2022 shows that x_rio achieves even better accuracies than the state of the art VIO framework Stereo VINS (no loop closures). We highly outperform VINS regarding runtime as x_rio with a single radar sensor runs >125x and even with two radar sensors almost 90x faster than VINS!
Check out our paper for more details x-RIO: Radar Inertial Odometry with Multiple Radar Sensors and Yaw Aiding.
If you use x_rio or the provided datasets for your academic research, please cite our related paper:
@INPROCEEDINGS{DoerJGN2022,
author = {Doer, Christopher and Trommer, Gert F.},
year = {2022},
month = {02},
pages = {329-339},
title = {x-RIO: Radar Inertial Odometry with Multiple Radar Sensors and Yaw Aiding},
volume = {12},
journal = {Gyroscopy and Navigation}}
The image below compares x_rio (blue) against groundtruth (red) on the office_floor datasets which is part of the Multi Radar Inertial Datasets JGN 2022. This trajectory covers 425m including two loops of the approximate same path. The final position error is just 0.25m which corresponds to 0.1% of the trajectory length.
Results on Multi Radar Inertial Datasets JGN 2022
Results comparing vins and x_rio regarding different radar sensor setups. Using two radar sensors (left and right) or even three radar sensors (left, center, right) improves the accuracy.
| Algorithm | office_floor | lab_floor | basement | workshop | Mean |
|---|---|---|---|---|---|
| VINS Stereo (no loop closures) | 0.48 | 0.47 | 0.52 | 0.18 | 0.41 |
| x_rio_plain_left | 0.23 | 0.27 | 0.55 | 0.49 | 0.39 |
| x_rio_plain_right | 0.40 | 0.33 | 0.53 | 0.43 | 0.42 |
| x_rio_plain_center | 0.59 | 0.91 | 2.21 | 0.53 | 1.06 |
| x_rio_plain_left_right | 0.17 | 0.34 | 0.58 | 0.43 | 0.38 |
| x_rio_plain_left_center_right | 0.16 | 0.32 | 0.62 | 0.38 | 0.37 |
| Algorithm | office_floor | lab_floor | basement | workshop | Mean |
|---|---|---|---|---|---|
| VINS Stereo (no loop closures) | 0.48 | 0.47 | 0.52 | 0.18 | 0.41 |
| x_rio_baro_yaw_left | 0.14 | 0.28 | 0.51 | 0.22 | 0.29 |
| x_rio_baro_yaw_right | 0.20 | 0.26 | 0.24 | 0.31 | 0.25 |
| x_rio_baro_yaw_center | 0.34 | 0.39 | 0.50 | 0.40 | 0.41 |
| x_rio_baro_yaw_left_right | 0.18 | 0.18 | 0.37 | 0.18 | 0.23 |
| x_rio_baro_yaw_left_center_right | 0.13 | 0.22 | 0.40 | 0.19 | 0.24 |
Run x_rio on the demo indoor datasets in rosbag mode, start rviz and plots:
roslaunch x_rio x_rio_demo.launch type:=rosbag_node enable_rviz:=True enable_plot:=True
Run x_rio using only two radar sensors:
roslaunch x_rio x_rio_demo.launch type:=rosbag_node enable_rviz:=True enable_plot:=False n_radar:=2
Run in online mode:
roslaunch x_rio x_rio_demo.launch type:=ros_node enable_rviz:=True enable_plot:=False
rosbag play --clock demo_x_rio.bag
Run simulation for a setup of three radar sensors and 10 Monte Carlo runs:
rosrun x_rio run_sim.py
Run and evaluate all runs of the Multi Radar Inertial Datasets JGN 2022 which produces the results shown in the two tables above:
rosrun x_rio evaluate_jgn2022_datasets.py _rosbag_base_dir:=<path_to_jgn2022_datasets> _eval_name:=<eval_name>
This will save the evaluation at <path_to_jgn2022_datasets>/results for all datasets. The estimation result of x_rio without and with barometer and yaw aiding for different radar sensor combinations are aligned with the ground truth and evaluated using rpg_trajectory_evaluation. Plots and further error metrics are located at <path_to_jgn2022_datasets>/results/evaluation_full_align.
Our x_rio implementation provides two nodes for ROS interfacing:
- ros_node: Subscribes to all topics and does online processing
- rosbag_node: Reads a rosbag and runs x_rio at maximum speed
A set of demo parameters is given in x_rio_demo_params. Most of the parameters can be changed online using rqt_reconfigure. Further documentation of the parameters can be found using the tooltip text in rqt_reconfigure and in the python files, see cfg.
Published topics are the same for both modes:
- ~state (rio/XRioState): full filter state (NED-convention)
- ~covariance (rio/XRioCovariance): diagonal elements of covariance matrix
- ~pose (geometry_msgs/PoseStamped): pose (ROS convention)
- ~twist (geometry_msgs/TwistStamped): twist (ROS convention)
- ~pose_path (nav_msgs/Path): pose path (ROS convention)
- ~combined_scan (sensor_msgs/PointCloud2): most recent scans of each radar sensor combined into a single scan given in the body frame
- ~yaw_aiding_init (sensor_msgs/PointCloud2): all radar scan point used for Manhattan angle init
- ~radar_scan_yaw_inlier (sensor_msgs/PointCloud2): radar scan point used for yaw aiding