A camera can tell an autonomous car that a lane line is roughly there. Getting from roughly there to a centimeter-accurate answer is the harder half of the problem, and it is the half a newly published NVIDIA application, US20260187862A1 — “Feature Location Identification,” published July 2, 2026 — sets out to address. The disclosed approach starts from a camera-derived guess at where a road marking sits, then uses LiDAR to sharpen that guess before the marking is handed to the planning and control stack. It is a sensor-fusion filing aimed at one narrow but load-bearing task: knowing precisely where the paint is.

In various examples, determining locations of road markings or other features for autonomous and semi-autonomous systems and applications is described. Systems and methods are disclosed that use LiDAR data—and/or other sensor data types—to determine the locations of road markings, such as lane lines, within environments. For instance, sensor data, such as image data, may initially be processed in order to determine an initial location associated with a road marking. The LiDAR data may then be used to increase the precision of the initial location associated with the road marking. For example, a LiDAR image(s) may be used to determine the actual location of the road marking, such as the center of paint associated with a lane line. Additionally, the initial location of the road marking may be updated, such as adjusted within a direction, based on the actual location of the road marking.— Feature Location Identification, US20260187862A1

The problem: cameras see the line, LiDAR measures it

Lane detection from images is mature, but a monocular or even stereo camera infers geometry rather than measuring it. Paint interpreted from pixels can drift with lighting, glare, wear, and the projection math that maps image space onto the road plane. The application frames the fix as a two-stage estimate. First, image data is processed to determine an initial location associated with a road marking. Then LiDAR data — described in the record as points within the environment, each carrying an intensity value — is used to determine a second location, and the initial location is updated to that second location. The engineering intuition is that lane paint is retroreflective: it returns LiDAR energy at a different intensity than the surrounding asphalt, so the point cloud contains a direct physical signature of the paint that the camera can only guess at.

The disclosed method narrows the search rather than scanning the whole scene. The abstract and claims describe using the camera-derived initial location to select an area of the environment associated with the road marking, and then pulling only the portion of LiDAR points that falls in that area. Within that portion, the intensity values are used to fix the actual location — the record refers to the center of paint associated with a lane line. One variant expresses both estimates as top-down images: a first top-down image carrying the camera’s initial location and a second top-down image generated from the LiDAR points, with the refined location determined by comparing the two. Casting both sensors into a common bird’s-eye frame is a standard way to make an image-space detection and a point-cloud measurement directly comparable.

Where it sits in the stack

The filing is explicit that this is a front-end perception step feeding the rest of autonomy: the refined marking location is used to perform one or more planning, navigation, or control operations. In other words, the output is not a prettier map but a more trustworthy input to where the vehicle decides to steer. That places the disclosure in the well-worn camera-versus-LiDAR debate on the side of fusion rather than either extreme — it presumes both sensors are present and assigns each the job it is physically better at: the camera for wide, cheap classification of what a marking is, the LiDAR for precise measurement of where it is. The record generalizes beyond lane lines, describing the same initial-then-refine pattern for “road markings or other features” and, in a broader independent claim, for an object located within an environment using depth data generally rather than LiDAR specifically.

The same July 2 drop shows the surrounding machinery this kind of perception module leans on. A companion application, US20260187981A1 (“Generating Training Data”), describes iteratively mining images with group classifiers to build cleaner positive and negative training sets for autonomous and semi-autonomous systems — the data-engine side of keeping a detector like this one accurate. Another, US20260187971A1, covers a multi-pass technique for removing artifacts from images in streaming pipelines, distinguishing true color edges from compression artifacts; edge fidelity is exactly what an image-space lane detector depends on before LiDAR ever refines it. Read together, the three describe a perception pipeline: generate the training data, clean the imagery, then fuse camera and depth to localize the feature.

Two honest caveats. A published application describes an approach and its claimed variants, not a shipped feature or a benchmark; nothing here reports how much accuracy the fusion buys or on what hardware. And the independent claims are notably device-shaped — they recite an autonomous machine with CPUs, GPUs, hardware accelerators, image sensors, and LiDAR or depth sensors, and a system-on-a-chip variant — which reads as much like a description of where the method runs as how it works. The drop also makes clear the disclosure is one thread in a much wider portfolio: the same publication run includes chip-design tooling such as US20260187332A1 for generative circuit layout and manufacturing methods like US20260186472A1 for die pairing. But for a perception reader, the interesting part is the specific bet: that the most reliable way to know where a lane line is is to let the camera propose and the LiDAR dispose.