US12674866B2, titled "Detection and classification of traffic signs using camera-radar fusion" and assigned to Waymo LLC, is a granted United States patent issued on July 7, 2026. It carries kind code B2, is reachable at its US12674866B2 record, and is classified under Cooperative Patent Classification codes G01S 7/417 (radar return processing), B60W 60/001 (autonomous driving control), and G01S 13/867 (radar sensor-fusion), among others. This brief reads what the record actually claims, in plain terms, rather than assessing its scope or worth.

Independent claim 1 is directed to a system rather than a method, and it names three cooperating parts of a vehicle: a sensing system, a data processing system, and a driving control system. The sensing system is configured to acquire two aligned inputs — a set of camera images of an environment and a set of radar images of the same environment. Everything downstream in the claim operates on those two streams together.

Walking the claimed pipeline

The heart of the claim is a sequence of neural networks. A first neural network generates a set of camera features characterizing the camera images. Critically, the claim specifies the granularity of those features: each camera feature is generated for a respective pixel of a coordinate system and for a respective time drawn from a plurality of times. A second neural network then generates radar features from the radar images, and the claim ties them to the same addressing scheme — each radar feature is generated for that same respective pixel and that same respective time. The two feature sets are therefore expressed on a shared per-pixel, per-time grid before any combination happens.

The claim then recites the fusion step in specific terms. The system fuses the camera feature set with the radar feature set to obtain a fused tensor, and the claim states that the fused tensor aggregates the two feature sets both (i) across pixels of the coordinate system and (ii) across multiple times of the plurality of times. In other words, the claimed representation is a single tensor spanning space and time, holding camera and radar information in common coordinates. A third neural network processes that fused tensor, and its output comprises the semantic content of one or more traffic signs in the environment. The final element of the claim closes the loop to actuation: a driving control system controls the vehicle based on that semantic content.

The abstract summarizes the same arrangement at a higher level:

The disclosed systems and techniques facilitate efficient detection and classification of traffic signs in driving environments. The disclosed techniques include, obtaining, using a sensing system of a vehicle a first set of perspective camera images of an environment and a second set of radar images of the environment. The techniques further include generating, using a first neural network, one or more camera features characterizing the first set of images, generating, using a second neural network, one or more radar features characterizing the second set of images, and processing the one or more camera features and the one or more radar features to obtain an identification of one or more traffic signs in the environment.— Detection and classification of traffic signs using camera-radar fusion, US12674866B2

Beyond the independent claim, the record discloses additional structure that colors how the pipeline is meant to run. Camera features are described as being mapped from a perspective coordinate system to a ground-surface coordinate system, which is what allows camera and radar features to share the pixel addressing recited in the claim. The third neural network is described as a backbone with classification heads, and the semantic content it produces is broken out into a sign's type, its value, and its relevance to the vehicle, together with the sign's location. A fourth neural network is disclosed as producing an auxiliary sign identifier that is fed back into the third network, the three networks are described as trained together, and the system is disclosed as identifying and eliminating duplicate signs.

It is worth isolating what the shared coordinate scheme does inside the claim, because it is the element that distinguishes an early fusion from a late combination of separate detectors. By requiring that each camera feature and each radar feature be generated for the same respective pixel and the same respective time, the claim keeps both modalities addressable on one grid, so the recited fused tensor is a joint representation rather than a merger of two independent outputs. The record frames the camera path in perspective image space and then maps it to a ground-surface coordinate system, which is the disclosed mechanism for bringing camera features into the same frame the radar features occupy. The neutral reading is that the claim recites where and when each feature lives before fusion, and that the fusion then aggregates across both of those axes.

Where it sits in the assignee's granted cohort

The patent issued alongside a set of related grants that together map onto a perception-and-planning stack. Also issued July 7, 2026 is US12675887B2, directed to high-throughput point cloud processing using a temporal LiDAR point-cloud network with time-point queries — a sensing counterpart to the sign-reading claim, operating on LiDAR rather than the camera-radar pair. Several further grants issued June 30, 2026 address the planning side. US12668282B2 is directed to trajectory prediction using diffusion models to predict multi-agent future trajectories, and US12669341B2 covers speed reasoning in the presence of uncertainty when pulling over, including reasoning about occluded parking spots.

Two more grants in the cohort touch sensing integrity and safety margins. US12671793B2 is directed to adjusting a vehicle sensor's field-of-view volume upon detected degradation, and US12668277B1 (kind code B1) claims using collision costs to maintain safe distances for autonomous vehicles. Read together with the hero record, the cohort is a set of issued patents spanning sensing, sign semantics, prediction, and control.

For readers tracking the record precisely: US12674866B2 is a granted patent, not a pending application. Its inventors are listed as Daniel Ho, Alper Ayvaci, Vasiliy Igorevich Karasev, and Jiakai Zhang. The claim language quoted and paraphrased here is drawn from the issued specification and claims as they appear in the patent record.