US20260177400
2026-06-25
Physics
G01C21/3841
The patent application describes an innovative system for generating data, creating maps from that data, and localizing vehicles to those maps. The system utilizes mapstreams, which are streams of sensor data, perception outputs from deep neural networks, and relative trajectory data, gathered from multiple vehicles over various drives. This data is then uploaded to the cloud for processing. The resulting outputs are high-definition (HD) maps that represent a comprehensive view, compiled from numerous data sources. Localization is achieved by comparing real-time sensor data to map data, allowing for precise positioning.
Autonomous vehicles rely heavily on mapping and localization for effective navigation. Traditional methods use expensive survey vehicles to create HD maps, limiting map availability to certain regions. These maps often become outdated due to changes in road conditions, and the update process is slow and resource-intensive. Furthermore, consumer vehicles typically lack the high-end sensors needed for precise localization, relying instead on less accurate GNSS data. This inaccuracy poses significant challenges for achieving reliable autonomous driving.
The proposed system addresses these challenges by enabling map creation and localization using consumer-grade sensors. Data is crowdsourced from numerous vehicles, reducing dependency on expensive equipment. Mapstreams are optimized by filtering unnecessary data and compressing the remaining information. This approach allows sensor data to be used directly for localization, enhancing accuracy. The system tracks relative trajectory information, enabling highly precise localization within centimeter-level accuracy.
Mapstreams are processed to create a fused HD map, integrating data from multiple drives to enhance robustness. As new data is collected, it is merged with existing information to update the map. Drive segments are geometrically aligned to optimize localization accuracy. The final fused HD map enables direct localization by comparing current sensor data with map data. Multiple sensor modalities can be used, with results fused to determine a precise localization outcome for each frame.
The system is versatile, applicable not only to autonomous vehicles but also to other domains such as robotics, aerial systems, and marine navigation. It can be implemented in various environments, including simulation settings. The system's functionality is supported by hardware, firmware, and software, with processes executed by processors running instructions stored in memory. This flexibility ensures the system's adaptability to different technological needs and environments.