HD Map & Localization Dataset V1.0 (Sangam)
NAVER LABS' open data set keeps being updated. We hope to grow together with researchers by sharing our latest data. On this page, you can check the composition and detailed information of the individual dataset 'HD map & Localization Dataset V1.0 (Sangam)'.
.gif)
Data Description
As part of the C-ITS demonstration project, this dataset consists of HD Map data of Sangam, Seoul, an autonomous driving pilot district and three data for location estimation obtained by scanning the HD Map area with the our autonomous vehicle, R1, and through this, localization performance for autonomous vehicle driving can be verified.
NAVER LABS' HD Map data is largely composed of three types of data - data-Road Layout (road vector data), LiDAR Feature Data (segmented point cloud), and Visual Feature Data (3D visual feature descriptor). Please note that LiDAR feature data is not included due to a legal issue.
Key Features
NAVER LABS HD Map & Localization Dataset is an efficient mapping method and has high positioning accuracy.
The hybrid HD mapping solution using aerial photographs and MMS equipment not only improves the trajectory position accuracy of MMS equipment, but also precisely matches various sensor data through Multi-sensor Calibration.
In addition, for high-accuracy localization, it is important to select and process information about roads and environmental structures well among many data obtained from the environment as well as highly accurate trajectory positioning. Our HD Map data contains road lanes and road surface location and geometry information in Road Layout, and structural information of road surroundings in LiDAR Feature Data and Visual Feature Data. Precise localization is possible through matching between sensor data.
Mapping Device : R1
R1 is a mobile mapping system developed for HD mapping. It includes various sensors such as cameras, 2D/3D LiDAR's, GPS, IMU, FOG, and wheel encoders, but is lighter than conventional MMS equipment in terms of cost. After automatically extracting features from the 2D/3D data collected by R1, we complete our HD map by summing the road layout information acquired from the aerial photograph taken in advance.
.png)
HD map Data Format
1. Road Layout
In the true-ortho photo generated through aerial image 3D Modeling (Bundle Adjustment/DSM/DEM, etc.), the following 7 detailed information such as road lane and road surface information are converted into Polyline, Polygon-type Vector data (shx, shp,dbf,prj)
- Lane
- Stop line
- Lane center line
- Guide line
- Road marking (driving direction and prohibiting driving)
- Road marking (surface type)
- Driving route node
2. Visual Feature Data
- Visual Feature Map is created by combining visual information from MMS camera images and 3D geometric information from MMS LiDARs, and is used as map data for image-based posture recognition (Visual Localization).
- Visual Feature Map consists of tile HDF5 (*.h5) files that are saved by dividing the surface into a grid structure, and the physical size of each tile is 20.48m x 20.48m.
- These tile files are stored in the same path as 15/{ty}/{tx}.h5, and the (tx=0,ty=0) tile is (32000.00 ~ 32020.48 m, 3662000.00 ~ 3662020.48 m) in the UTM coordinate system.
- Each tile h5 file is composed of 4 attributes of {points, rpy, desc, pose}, and each has the following meaning. (N is the number of visual keypoints in the tile, it can be different for each tile)
.jpeg)
3. LiDAR Feature Data (not yet disclosed)
- Precise data in the form of LiDAR point cloud is not included in this open data set as it cannot be disclosed to an unspecified number of issues due to legal issues.
- Please refer to it as a component of the LiDAR feature data of NAVER LABS HD map.
- The map created with the point cloud acquired using the LiDAR sensors is saved in las file format version 1.2.
- The point cloud stored in the las file is saved in the standard point format 0 type of the las file.
- Each point has fields of x, y, z, intensity, and class label.
- Intensity is saved as a value ranging from 0 to 65535 after normalizing.
- The number stored in the class represents the semantic label of an individual point. Dynamic objects such as vehicles on the road have been removed. The meaning of the label number is as follows:
Localization Data Format
Localization data includes raw sensor data of lidar (6), camera (4), GPS, IMU, OBD, FOG, and Wheel Encoder. It also includes sensor calibration data for data transmission between multiple sensors and pseudo GT pose information accurately estimated through mapping for the purpose of evaluating the localization algorithm developed.
