Autoware Labs urban dataset mapping with MAP IV Engine

[n-patiphon]

Hi all,

I would like to share some of our recent activities at MAP IV, Inc. in the Localization and Mapping working group.
Recently Leo Drive collected and published a dataset that contains various types of environments that are typically challenging for mapping and localization tasks, such as tunnels and bridges. (see here for more details).
With this dataset, we have used our mapping software (MAP IV Engine) to create geo-referenced noise-free point cloud maps.

First, we tried creating the maps using only LiDAR data to test the quality of the point cloud. We could complete the mapping of all five sections included in the dataset. The results are also satisfactory as the details of the map seem clear.

Mapping results with LiDAR only

Then, we proceeded to create the map again with other sensor data, i.e., GNSS, IMU, and Odometry. Since the sensor configurations are quite different from our mapping hardware, SEAMS, which MAP IV Engine is optimized for, we needed to do some data conversions. With GNSS, we could create geo-referenced point cloud maps.

The overall trajectory of Section 1 to Section 4 (Section 5 doesn’t have GNSS data)

Mapping results in some challenging locations (LiDAR, GNSS, IMU, and Odometry)

Although we could achieve good mapping results in most of the challenging locations, it is important to note that introducing other sensors can degrade the mapping results if the outputs from those sensors are not reliable.

These are the comparisons between GNSS (Red + Blue) and MAP IV Engine (Green) trajectories.

As can be seen from the figures above, the GNSS trajectories are likely incorrect as there are abrupt changes in lateral positions. These abrupt changes affect the final mapping outputs.

These are the mapping results where GNSS was inaccurate.

Our mapping software, MAP IV Engine, doesn’t actually need highly accurate GNSS data. In fact, it works well with low-cost GNSS as long as you can obtain the raw GNSS data. The GNSS data, as well as IMU and odometry, included in this dataset, on the other hand, has already been processed. This poses challenges when trying to incorporate these data into the mapping process.

Once geo-referenced maps were created, we then applied the automatic noise removal feature of MAP IV Engine to remove dynamic objects in the point cloud maps. This step is very crucial especially if you collect and create maps in busy urban areas as it is impossible to avoid all other vehicles and pedestrians.

As can be seen in the raw point cloud, there are traces of moving vehicles and several parked vehicles in the map.

After applying automatic noise removal, we could detect and remove these undesired objects and obtain a cleaner map.

Another example of raw and noise-removed point cloud

To sum our report up a little bit, we used our mapping software, MAP IV Engine, to create noise-free point cloud maps from the dataset collected by Leo Drive. Although the dataset contains several locations where mapping and localization are typically prone to fail, such as tunnels and bridges, we could create the point cloud maps of the entire route. It is important to note that in locations where GNSS data is unreliable, the mapping results may not be accurate.

There are a few things that could be done to improve the mapping results. First, instead of collecting data separately in multiple sections, it would be better to collect all the data in a single session and then split it later if needed. This will eliminate any inconsistency between each scene. Also, some parameter fine-tuning or modifications of our mapping software to make use of GNSS reliability output by Applanix POS LVX used to capture the data will likely improve the results. However, it was our intention to test our mapping software as is, and since MAP IV Engine is optimized to work best with our hardware, it was expected that we would not achieve the same level of accuracy as we would have had the data been collected by our hardware, SEAMS, especially SEAMS LX which is our mapping hardware designed to be mounted on a vehicle.

We will later share the PCD files of the maps so that they can be used for Autoware testing and development by the community (please feel free to contact us and mention this discussion if you would like to use or create new maps for commercial use).

[n-patiphon]

Here's the PCD files.

• Section 1 (map2) : https://drive.google.com/drive/folders/1KB9XdKFV3qdYNqinrdWUJno9dgGkrXZk?usp=sharing
• Section 2 (map3): https://drive.google.com/drive/folders/1ZY0ptSKJDtmRxiAzbWpt5pm1LR77rR4n?usp=sharing
• Section 3 (map4): https://drive.google.com/drive/folders/120pXigKmcwKt35bizfzDTuSOIiySn6if?usp=sharing
• Section 4 (map5): https://drive.google.com/drive/folders/1m5-HcErMYc4Foj-sN7hEZYh8_g5U3wD7?usp=sharing
• Section 5 (map6 no gnss): https://drive.google.com/drive/folders/14PdIm-XBrIthj-5ta-fdPaTZOQsyyjKF?usp=sharing

Each section contains clean PCD and noise PCD files, along with the trajectory in both KML (Lat, Lon) and CSV (MGRS). All PCD files except Section 5 are in MGRS coordinates (35TPF).