KLTDataset

The KLT dataset is a light urban scenario dataset containing GNSS raw measurements collected using a U-blox F9P receiver, ground truth provided by the SPAN-CPT system, and LOS/NLOS labels for GPS and Beidou satellites. Additional data, such as IMU, LiDAR, and camera recordings, are also included in the ROS bag file.

We also provide a start script and configuration files to enable researchers to get started with the dataset quickly and efficiently.

Quick Start

1. Clone the repository

   git clone https://github.com/ebhrz/KLTDataset

2. Install the dependencies

   cd KLTDataset
   pip install -r requirements.txt

3. Download the data from Dropbox, and extract data and label to the repository folder. The folder tree should looks like:

   .
    ├── config
    │   ├── 0610_klt1_203.json
    │   ├── ...
    ├── data
    │   ├── GNSS
    │   └── ROS
    ├── label
    │   ├── 0610_KLT1_203
    │   ├── 0610_KLT2_209
    │   ├── 0610_KLT3_404
    │   ├── 1109_KLT1_421
    │   ├── 1109_KLT2_294
    │   ├── 1109_KLT2_301
    │   ├── 1109_KLT3_666
    │   ├── 1116_KLT1_184
    │   ├── 1116_KLT2_169
    │   └── 1116_KLT3_649
    ├── LICENSE
    ├── README.md
    ├── requirements.txt
    ├── rtk_util.py
    └── demo.py
   

   If you don't use the ROS data, you can only download the label and data/GNSS folder, because the bag file is huge.
4. Now, run the demostration.

   python demo.py config/0610_klt1_203.json

   or

   python demo_rtklib.py config/0610_klt1_203.json

   The first script will utilize my SPP implementation for positioning. The second script will employ both pntpos and rtkpos from RTKLIB for positioning purposes. Then you can see the figure.

Now start to explore the dataset via the demo python script! If you find our dataset useful, we would greatly appreciate it if you could cite our work:

@inproceedings{hu2023fisheye,
  title={Fisheye camera aided NLOS exclusion and learning-based pseudorange correction},
  author={Hu, Runzhi and Wen, Weisong and Hsu, Li-ta},
  booktitle={2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC)},
  pages={6088--6095},
  year={2023},
  organization={IEEE}
}

Introduction

Datasets are essential for training machine learning (ML) and deep learning (DL) models, especially when it comes to accurate labeling, which is critical for NLOS classification and pseudorange correction. Although several urban canyon datasets offer raw GNSS measurements and ground truth, only a limited number provide labeled NLOS data. This gap often forces researchers to rely on proprietary datasets, leading to a lack of standardized benchmarks. Widely recognized datasets such as KITTI¹, ApolloScape², Waymo³, and nuScenes⁴ omit raw GNSS data altogether. Our open-source datasets, UrbanLoco⁵ and UrbanNav⁶, do include raw GNSS measurements and positional ground truth but do not feature NLOS labels. Mytuc⁷ takes a step towards identifying NLOS satellites by comparing them to ground truth, yet this approach lacks sufficient precision. In previous work⁸, we used a 3D map to classify NLOS satellites, but pre-built maps don't account for real-time obstructions such as trees or vehicles.

In conclusion, the absence of a dataset that offers both raw GNSS measurements and accurate NLOS labels continues to hinder researchers seeking comprehensive benchmarks. To address this gap, we leveraged our proven equipment platform and data collection methods⁶ to develop the Kowloon Tong (KLT) dataset. This dataset accounts for temporal variations, satellite distribution, and differing lighting conditions. Through our automated labeling technique⁹, we achieve a high degree of accuracy (approximately 99%) in identifying NLOS satellites. A detailed comparison of this dataset with others is provided in the table below.

Existing Dataset	Environments	Raw GNSS Data	Ground Truth	IMU	LiDAR	Fisheye Camera	NLOS Label
KITTI1	Urban	❌	✅	✅	✅	❌	❌
ApolloScape2	Urban	❌	✅	✅	✅	❌	❌
Waymo3	Urban	❌	✅	✅	✅	❌	❌
nuScenes4	Urban	❌	✅	✅	✅	❌	❌
UrbanLoco5	Urban	✅	✅	✅	✅	❌	❌
UrbanNav6	Urban	✅	✅	✅	✅	❌	❌
Mytuc7	Urban	✅	✅	❌	❌	❌	✅
KLT (Proposed)	Urban	✅	✅	✅	✅	✅	✅

The performance of our automated labeling technique is shown in below table:

Dataset	T/F	NLOS(P)/LOS(N)	Accuracy	Precision	Recall
KLT1-203	T F	1408/1859 13/5	99.45%	99.09%	99.65%
KLT2-209	T F	1706/1747 15/6	99.37%	99.12%	99.66%
KLT3-404	T F	2262/3909 18/9	99.53%	99.23%	99.59%

Collection Platform

Below figure shows our platform. For more details, please refer to our UrbanNav Dataset

Summary Of The Dataset

The below table shows the summary of the dataset. The dataset includes 10 subsets listed in the table, which vary in time, satellite distribution and illumination. Please noticed that we only labeled GPS and Beidou satellites. And because the illumination condition is very low in KLT3-649, so the label is not reliable.

Dataset No.	Date	Time	Area	Illumination	Epoch	Samples	LOS/NLOS
KLT1-203	2021.06.10	11:35-11:38	KLT1	High	203	3295	1862/1433
KLT2-209	2021.06.10	11:48-11:51	KLT2	High	209	3498	1738/1760
KLT3-404	2021.06.10	11:52-11:58	KLT3	High	404	6309	3918/2391
KLT2-301	2023.11.09	17:37-17:42	KLT2	High	301	4838	3154/1684
KLT2-294	2023.11.09	17:43-17:48	KLT2	High	294	4992	3253/1739
KLT1-421	2023.11.09	17:05-17:12	KLT1	High	421	6607	4197/2410
KLT3-666	2023.11.09	17:12-17:23	KLT3	Medium	666	10085	6481/3604
KLT2-169	2023.11.16	17:50-17:53	KLT2	Medium	169	2713	1710/1003
KLT1-184	2023.11.16	17:53-17:56	KLT1	Low	184	2891	1877/1014
KLT3-649	2023.11.16	17:56-18:07	KLT3	Very Low	649	9826	/

Data

For more information on how to use the dataset, you can refer to the demo.py.

GNSS Data

GNSS data is collected by U-blox F9P. The observation file is stored in data/GNSS/*/COM*.obs. Please notice that we collect data from four constellations including GPS, Beidou, Galileo and BLONASS, but only GPS and Beidou satellites are labeld.

Reference Station Data

We are sorry that we can not directly provide the reference station data due to the policy of Hong Kong Lands Department. You can download the reference station data from here. We recommand you use the data from HKSC station.

Ground Truth

Ground truth is stored in label/*/gt.csv. The first column is the timestamp(in GPS time!!! convert to UTC need to add the leap seconds 18), and the rest are the LLH coordinate.

Label Data

Label data is in label/*/nlos.pkl. This file is generated by python pickle. You can use follow script to read the content:

import pickle
import pyrtklib as prl

def satindex2name(sats):
    name = prl.Arr1Dchar(4)
    if not isinstance(sats,list):
        prl.satno2id(sats+1,name)
        return name.ptr
    names = []
    for i in sats:
        prl.satno2id(i+1,name)
        names.append(name.ptr)
    return names

with open("label/0610_KLT2_209/nlos.pkl",'rb') as f: # KLT1_203 for example
   labels = pickle.load(f)

for label in labels:
   print("the timestamp: ", label[0]) # timestamp
   print("all the satellite index: ", label[1]) # start from 0, please +1 to get the actual id
   print("all the satellite name: ", satindex2name(label[1]))
   print("LOS satellite index: ", label[2]) # the LOS satellites index
   print("LOS satellite name: ", satindex2name(label[2]))
   print("the corresponding flatten fisheye image shape: ", label[3].shape)
   for satid in label[4]:
      part = label[4][satid]
      print(f"satellite {satid}: the part image shape of the satellite projection on the flatten fisheye image: ", part[0].shape)
      print(f"satellite {satid}: the coordinate of satellite the projection on the flatten fisheye image ({part[1]/3.46875},{part[2]/3.46875})")

Image Data

Also in the label folder, there is "image.zip" file. There are three kinds of images in the zip file, including original fisheye image with satellite projection, flatten fisheye image with satellite projection, and the sky segmentation of the flatten fisheye image. There are examples below:

ROS Bag

For data in 2021/06/10, the camera topics are /zed2/camera/left/image_raw and /zed2/camera/right/image_raw. For data in 2023/11/09 and 2023/11/16, the camera topic is /pointgrey/left_camera/image_color/compressed. For all the bag files, the fisheye camera topic is /fisheye/image_color/compressed, the Velodyne 32 HDL topic is /velodyne_points. Two 16-scan LiDARs' topic are /left/lslidar_point_cloud and /right/velodyne_points. The IMU data is at /imu/data. For the extrinsic parameters, please refer to UrbanNav Dataset

Trajectory

KLT1-203	KLT2-209

KLT3-404	KLT1-421

KLT2-294	KLT2-301

KLT3-666	KLT1-184

KLT2-169	KLT3-649

Creators

Dr. Li-Ta Hsu, Dr. Weisong Wen, Runzhi Hu, Feng Huang, Yihan Zhong, Xiao Xia, and Ivan Ng were instrumental in creating this dataset.

Reference

Footnotes

1. A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? The KITTI vision benchmark suite,” in 2012 IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 2012, pp. 3354–3361. ↩ ↩²

2. X. Huang, X. Cheng, Q. Geng, B. Cao, D. Zhou, P. Wang, Y. Lin, and R. Yang, “The ApolloScape dataset for autonomous driving,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018, pp. 954–960. ↩ ↩²

3. P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V. Patnaik, P. Tsui, J. Guo, Y. Zhou, Y. Chai, B. Caine et al., “Scalability in perception for autonomous driving: Waymo open dataset,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 2446–2454. ↩ ↩²

4. H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, “nuScenes: A multimodal dataset for autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 11 621–11 631. ↩ ↩²

5. W. Wen, Y. Zhou, G. Zhang, S. Fahandezh-Saadi, X. Bai, W. Zhan, M. Tomizuka, and L.-T. Hsu, “UrbanLoco: A full sensor suite dataset for mapping and localization in urban scenes,” in 2020 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2020, pp. 2310–2316. ↩ ↩²

6. L.-T. Hsu, F. Huang, H.-F. Ng, G. Zhang, Y. Zhong, X. Bai, and W. Wen, “Hong Kong UrbanNav: An open-source multisensory dataset for benchmarking urban navigation algorithms,” NAVIGATION: Journal of the Institute of Navigation, vol. 70, no. 4, 2023. ↩ ↩² ↩³

7. P. Reisdorf, T. Pfeifer, J. Breßler, S. Bauer, P. Weissig, S. Lange, G. Wanielik, and P. Protzel, “The problem of comparable GNSS results–An approach for a uniform dataset with low-cost and reference data,” in Proc. of Intl. Conf. on Advances in Vehicular Systems, Technologies and Applications (VEHICULAR), 2016. ↩ ↩²

8. G. Zhang, P. Xu, H. Xu, and L.-T. Hsu, "Prediction on the Urban GNSS Measurement Uncertainty Based on Deep Learning Networks with Long Short-Term Memory," in IEEE Sensors Journal, vol. 21, no. 18, IEEE, 2021, pp. 20563–20577. ↩

9. R. Hu, W. Wen, and L.-T. Hsu, “Fisheye camera aided GNSS NLOS detection and learning-based pseudorange bias correction for intelligent vehicles in urban canyons,” in 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC), IEEE, 2023, pp. 6088–6095. ↩