The arms dealer of autonomy has a recurring move: make every sensor look like something its chips already process well. US11906660B2, granted to Nvidia in February 2024, does it for lidar — "Object detection and classification using LiDAR range images."
Classified under G01S 17/931 (vehicle lidar), G06V 10/82 (neural-network image processing) and G06V 20/58 (object detection), the patent's trick is representation. A lidar produces a 3D point cloud, which is awkward for the convolutional networks that dominate image AI. By projecting that cloud into a 2D 'range image' — each pixel encoding distance — the data becomes something those proven, efficient vision networks can chew on directly.
Here is the engineering payoff. The deep-learning toolchain is overwhelmingly built for images: efficient convolutions, mature training recipes, hardware tuned for 2D feature maps. Reshaping lidar into an image-like form lets a lidar perception stack ride that entire ecosystem instead of inventing native 3D machinery from scratch. It is reuse as strategy.
The honest trade is information geometry. Projecting 3D points to a 2D range image flattens some spatial relationships and can introduce projection artifacts — two points far apart in 3D may land near each other in the image. The patent's value is in handling that distortion well enough that the efficiency gain outweighs the geometric compromise.
This is Nvidia being Nvidia: it is not picking a side in the sensor wars, it is making sure that whichever sensors win, the perception runs efficiently on its silicon. A range-image approach to lidar slots neatly into the same hardware that accelerates camera networks — one platform, many sensors.
For readers watching who profits from autonomy regardless of who wins it, this patent is the pattern. The robotaxi operators argue about sensors; Nvidia quietly patents the representations that make every sensor's data run fast on its chips. The range image is one more brick in that road.