The Data That Paves the Way: How We’re Building the First Open Dataset for HD Maps

We are building the first open dataset for maintaining and updating HD maps with Zenuity, AstaZero, RISE, and AI Innovation of Sweden. Together, we will collect map data in a highly controlled environment through cheap dashcams, lidar, and radar, in an effort to build a cost-effective way of updating HD maps and teaching autonomous vehicles to understand their surroundings through an HD map — even when the surroundings have changed remarkably.

Maps are crucial for autonomous vehicles to understand their surroundings. Self-driving cars need meticulously detailed and accurate data to inform them about everything from the presence of humans and other vehicles to turn restrictions and speed limits. Street scenes change constantly though, which puts pressure on vehicles to understand changes and what they actually mean. For instance, a speed limit sign that’s no longer visible because it has been covered by snow does not mean that the speed limit has changed or that it no longer applies.

That’s why we entering a research project with Zenuity, AstaZero, RISE, and AI Innovation of Sweden. Together, we will build the first open dataset for updating and maintaining HD maps. The research project will take place at AstaZero, the first full-scale independent test environment for autonomous vehicles, where an autonomous vehicle will collect data over the course of two years through cameras, lidar, radar, and GPS.

The AstaZero test track The AstaZero test track

The research and innovation centre RISE is placing trackable anchor points alongside the test tracks. In combination with data collection through cheap cameras, the anchor points allow the vehicle to detect and understand changes in the street scenes with high accuracy. Detecting and understanding change means that the vehicles will know what traffic rules apply where — even when the road markings and traffic signs are obscured. By using low-end cameras, we hope to make significant advances in self-driving technology relying more on imagery and computer vision as opposed to lidar. Lidar is great from a technology perspective, but it is unlikely that self-driving technology will be rolled out on a larger scale as long as it depends on lidar as much as it does today. It’s simply too expensive. Cameras and computer vision, on the other hand, are not.

The research project started in October this year and will run until 2022. It is partially funded by FFI of Vinnova, the Swedish innovation fund agency. As with both Mapillary Vistas Dataset and Mapillary Traffic Sign Dataset, this dataset will be available under a free research license to help advance future research.

/Emil, VP of Automotive

Continue the conversation