IRS Radar Thermal Visual Inertial Datasets IROS 2021

This dataset can be downloaded here and can be processed by our RRxIO pipeline.

Further details regarding the datasets can be found in our paper.

Contact me in case of questions or issues: Christopher Doer

The dataset consists of 5 datasets recorded at a motion capture lab and 4 datasets recorded in typical scenes indoors and outdoors. A short summary of the datasets is given below:

Motion Capture Datasets

Ground truth is available and was created using vicon2gt which does a batch optimization of the IMU and motion capture measurements to create the ground truth.

NameTrajectory LengthDurationEnvironment
MoCap easy36m94sMotion capture lab, low dynamic
MoCap easy83m87sMotion capture lab, medium dynamic
MoCap easy70m86sMotion capture lab, high dynamic
MoCap dark111m135sMotion capture lab, low dynamic, lights are switched of
MoCap dark fast75m86sMotion capture lab, medium dynamic, lights are switched of

Pseudo Ground Truth Datasets

Pseudo ground truth is available and was created using VINS with loops closure and subsequent manual removal of scale errors.

NameTrajectory LengthDurationEnvironment
Gym74m84sTwo loops within a gym hall
Indoor floor240m206sTwo loops in an office building
Outdoor campus102m102sOutdoors between two buildings
Outdoor street264m186sStreet setting with cars, buildings and vegetation

image

Cite

If you use this dataset for your academic research, please cite our related paper:

@INPROCEEDINGS{DoerIros2021,
  author={Doer, Christopher and Trommer, Gert F.},
  booktitle={2021 IEEE/RSJ International Conference on Intelligent Rotots and Sytems (IROS)}, 
  title={Radar Visual Inertial Odometry and Radar Thermal Inertial Odometry: Robust Navigation even in Challenging Visual Conditions}, 
  year={2021}}
}

Known Issues

Sensor Setup

The sensor platform is equipped with an FMCW radar sensor (TI IWR6843AOP), a thermal camera (FLIR Boson 640), a monochrome visual camera (IDS UI-3241) and an IMU (Analog Devices ADIS16448). Accurate temporal synchronization of all sensor data is achieved using a micro-controller board. The sensor rates are: IMU (409 Hz), visual and thermal cameras (20 Hz) and radar (10 Hz). The TI IWR6843AOP is a single chip 60 GHz Fast FMCW radar with an integrated antenna array (3 Tx and 4 Rx) and a Field of View (FOV) of approximately 120 deg in azimuth and elevation. All radar processing is done on chip. The visual camera uses a global shutter and provides images of 1280 × 1024 px. The thermal camera records with 640 × 512 px.

image

Calibration Files

The individual calibration files including the extrinsic transforms of the sensor rig:

Data Format

Each dataset consists of two files (details below):

Rosbag

The sensor data is provided with rosbags containing the following topics:

The point cloud point type is sketched below, for an example implementation see radar_point_cloud.h and radar_point_cloud.cpp:

struct RadarPointCloudType
{
  PCL_ADD_POINT4D;      // position in [m]
  float snr_db;         // CFAR cell to side noise ratio in [dB]
  float v_doppler_mps;  // Doppler velocity in [m/s]
  float noise_db;       // CFAR noise level of the side of the detected cell in [dB]
  float range;          // range in [m]
  EIGEN_MAKE_ALIGNED_OPERATOR_NEW
} EIGEN_ALIGN16;
                                  
pcl::PointCloud<RadarPointCloudType> your_pcl;

Pseudo Ground Truth

The pseudo ground truth is provided with csv files (separator: whitespace).
The format is: timestamp p^w_x p^w_y p^w_z q^w_x q^w_y q^w_z q^w_w
with the global position p^w and attitude quaternion q^w.
This file format is compatible e.g. with rpg_trajectory_evaluation for trajectory evaluation as done in our paper.

Evaluation Result of Doer et.al. IROS 2021

The evaluation result using the approach of our related paper is also included in this dataset. These results can be created using a single script resulting in the analysis below:

rosrun rrxio evaluate_iros_datasets.py <path-to-rtvi_datastets_iros_2021>

MoCap Datasets

image

Pseudo Ground Truth Datasets

image

Average Realtime Factors

image