NVIDIA has unveiled a platform that combines the accuracy of its DeepMap ground-truth maps with the modernity and scale of AI-powered crowdsourced maps.

NVIDIA unveiled NVIDIA DRIVE Map, a multimodal mapping platform with stability and autonomy, at its GTC 2022 keynote on the 4th.

Drive Maps plans to provide survey-quality ground maps of 500,000 kilometers of roads in North America, Europe, and Asia by 2024.

DriveMap includes multiple localization layers that can be used with camera, radar, and lidar modalities. DriveMap can independently identify location based on each map layer, providing the diversity and redundancy necessary for the highest level of autonomy. The camera location identification layer consists of map attributes such as lane markings, road markings, road boundaries, traffic lights, and poles.

▲ Drive Map Semantic Location Identification Layer

The radar positioning layer is a collective point cloud of radar returns. It is particularly useful in conditions with high camera loads, such as lighting conditions or adverse weather conditions that place a strain on the camera or lidar.

▲ Drive Map Radar Location Identification Layer

Even in rural areas where standard map attributes are indistinguishable, radar location identification is useful. Drive Map can pinpoint locations based on surrounding objects that generate radar returns. Lidar voxel layers provide the most accurate and stable representation of the environment. It also presents the world in 3D with a resolution of 5 cm, an accuracy that cannot be achieved with cameras and radar.

▲ Drive Map Lidar Voxel Location Identification Layer

Once you've identified your location on the map, Drive Maps can use the detailed semantic information the map provides to help you plan ahead and make safe driving decisions.

DriveMap is built using two mapping engines: a ground-based map engine and a crowdsourced map engine, to collect and maintain comprehensive information about the Earth's surface. This approach combines the strengths of both, achieving centimeter-level accuracy with professional surveying vehicles, while also achieving the up-to-dateness and scale that would require millions of passenger vehicles to continuously update and expand the map.

The ground truth engine is based on the DeepMap ground truth map engine, a technology developed and proven over the past six years. The AI-based crowdsourcing engine collects updated map information from millions of vehicles on the road, continuously uploading new data to the cloud. The data is then collected in Omniverse and used to update maps, which wirelessly update actual vehicles with the latest map information within hours.

Drive Map also supports passenger vehicles that meet Drive MapStream requirements to continuously update their maps using camera, radar, and lidar data through a data interface called Drive MapStream.

DriveMap accelerates autonomous vehicle (AV) development by supporting AI to make optimal driving decisions and generating ground truth training data for deep neural network (DNN) training, testing, and validation.

At its core is the Omniverse, where actual map data is loaded and stored. The Omniverse represents a digital twin, on a global scale, that is continuously updated and expanded by map surveying vehicles and millions of passenger vehicles.

Detailed maps, rendered through Omniverse's built-in automatic content generation tools, are transformed into a drivable simulation environment for use with NVIDIA DRIVE Sim. Topographical features such as road elevations, road markings, islands, traffic signals, signs, and vertical poles are precisely replicated with centimeter-level accuracy. Autonomous vehicle (AV) developers can use the simulation environment, leveraging physics-based sensor simulation and domain randomization, to generate training scenarios unavailable in real-world data.

Additionally, scenario creation tools can be used to test AV software in a digital twin environment before deploying autonomous vehicles in real-world environments. Furthermore, digital twins provide operators with a perfectly virtualized representation of the environment in which the vehicle will operate in the real world, enabling remote operation when necessary.