The edge case the patent quietly admits is a budget problem: a self-driving car cannot afford to reason about everything in its view equally. US11556127B2, granted to Baidu USA (the Apollo program) in January 2023, splits the world to save effort — "Static obstacle map based perception system."

Classified under G05D 1/0088 (autonomous control), G01S 17/931 (lidar) and G06T 7/246 (object tracking), the patent's move is to build and maintain a map of the static obstacles — parked cars, walls, barriers, curbs — and treat them separately from dynamic agents. The unmoving world gets logged once; the moving world gets the expensive, continuous prediction.

“The offline map generation process may collect multiple point cloud data of the same area.”— U.S. Patent No. 11,556,127 source

The claimed method is a two-phase pipeline, and the filtering logic in claim 1 is more careful than "static versus dynamic." The offline phase collects point-cloud data for an area — claim 2 specifies capturing the same area "for a plurality of repetitions," which is how you tell what actually stays put — then runs the perception algorithm to detect all obstacles. From that set it performs two successive removals. First it removes "one or more moving obstacles and one or more stationary obstacles that are movable," leaving genuine static obstacles; that second category is the subtle one — a parked car is stationary right now but movable, and the claim explicitly strips it out rather than trusting it. Second, it removes non-movable stationary obstacles that fall outside an area of interest, leaving a focused subset. That subset is recorded (claim 5: its position and shape) as the "predefined static obstacles" and baked into a map.

The online phase, in claim 6, is where the compute savings land. The running vehicle takes live sensor data and "trim[s] one or more regions-of-interest (ROI)… to exclude one or more of the predefined static obstacles," runs the perception algorithm only on the trimmed ROI to get a partial perception result, then adds the known static obstacles back in to complete the output. In other words: don't re-perceive the wall you already mapped; spend the cycles on the ROI where things can change, and re-attach the static furniture afterward. Claim 8 confirms the point of the trimmed pass is "detecting one or more moving objects in the ROI." It is a compute-allocation strategy disguised as a perception architecture — prediction is costly, and most of what a lidar sees is not going anywhere.

One dependent claim quietly addresses the obvious safety objection. If you subtract mapped static obstacles, do you risk forgetting one that still matters — a barrier right next to you? Claim 9 says the system adds back the excluded static objects "that are less than a threshold distance from the ADV." Distant mapped clutter can stay subtracted; nearby mapped structure is re-inserted into the perception output so the planner never loses track of a wall it is about to scrape. That threshold is the patent's hedge against its own optimization.

The CPC tags are not decoration; each names a stage of this pipeline. G01S 17/931 is lidar for vehicle perception — the point clouds claim 1 collects come from exactly that sensor class. G06T 7/246 is image-sequence object tracking with motion detection, which is what claim 3 invokes when it has the offline perception algorithm "detect[] motions of the plurality of obstacles": the way the system separates moving from movable-but-stationary from truly static is by watching how each return behaves across the repeated passes claim 2 demands. And G05D 1/0088 is the autonomous-control bucket where the resulting map is consumed — claim 1 ends by "generating a map… utilized by the perception algorithm of the ADV to navigate through the area." Read in order, the classification is a compressed description of the same offline-to-online flow the claims spell out: lidar in, motion-based sorting in the middle, a navigation map out. It is a small reminder that for an algorithm patent, the CPC codes often map more faithfully onto the method than the marketing language ever does.

Here is the pragmatism worth noting. By segregating static structure into a map, the system frees its real-time reasoning to focus on the pedestrians, cyclists and cars that can actually surprise it. The repetition requirement in claim 2 is doing real epistemic work: an obstacle that appears across multiple passes of the same area is far more likely to be true static structure than a one-off ghost return or a temporarily parked vehicle, and building the map from repeated point clouds is how the system earns the right to ignore it later.

The honest risk is the misclassification at the boundary. The strategy depends on correctly sorting static from dynamic, and the dangerous failure is the parked car that suddenly pulls out — an object the system filed as static that becomes dynamic. The patent's two-stage removal in claim 1 is precisely an attempt to defuse this: by separating "stationary but movable" from "static," it tries not to map away the very objects most likely to move. But a map is a snapshot, and the threshold-distance re-insertion of claim 9 is the runtime backstop for when a mapped object near the vehicle behaves like a live one. The patent's real burden is handling exactly those transitions without being fooled.

That Baidu's Apollo program holds this is a reminder that the autonomy race is global and that Chinese AV development files heavily in the U.S. patent system. Apollo's perception choices mirror the same trade-offs Waymo and Cruise face: how to make perception fast enough to be safe on a real compute budget. The claims also span the usual trio of forms — method (claim 1), non-transitory machine-readable medium (claim 10), and data-processing system (claim 19) — the standard way an AV filer fences an algorithm across software, storage, and hardware so a competitor cannot route around it by shipping the same logic in a different package.

For readers comparing AV stacks, the static-versus-dynamic split is a useful lens. A stack that treats the whole scene uniformly is either compute-rich or naive; one that explicitly separates the fixed world from the moving one — and, like this patent, keeps a distance threshold to re-admit nearby static structure — has confronted the budget reality that real-time driving imposes.