A published patent application is a roughly eighteen-month-delayed look at where a company was spending its research effort. So when a stereo-vision specialist publishes a filing that reaches past raw depth maps and into the occupancy grid — the structured, bird's-eye representation an autonomous driving stack uses to decide what space is free to move through — the filing is worth reading as a directional signal. That is what Nodar's US20260087824A1, published March 26, 2026, does.

The application, "Camera-Based Dynamic Occupancy Grid and Related System and Method," describes obtaining point-cloud data from camera sensors on a vehicle, compressing it to a bird's-eye-view representation, establishing a grid of cells, and tracking each cell's occupancy across successive frames with noise reduction before outputting an occupancy map to the vehicle's controller. Its classification, G06V 20/58 (recognizing objects relevant to vehicle operation) with G06T 7/20 (motion tracking), places it in the perception-output corner of the field. The point worth noting for a business reader is the word camera: occupancy grids are a representation that mapped-and-LiDAR autonomy stacks have historically produced from range sensors, and this filing describes deriving one from camera data.

For each cell, determining occupancy values of the cell for a time tn and performing noise reduction by assigning weights to the cell based on predicted occupancy values of the cell for the time tn and on occupancy values of the cell for a previous time tn−1.— Camera-Based Dynamic Occupancy Grid and Related System and Method, US20260087824A1

Where the cluster has been heading

The new filing reads differently once it is placed against the body of applications Nodar has published over several years. The company's foundational disclosures describe a "non-rigid stereo vision camera system" — US20230005184A1 and the more recent US20240242382A1 — using long camera baselines of roughly one to four meters and continuous tracking of camera parameters so that relative motion between cameras does not degrade the depth map, removing the need for factory calibration. That is the engineering premise of the whole portfolio: get long-range depth from widely spaced cameras without a rigid, pre-calibrated rig.

From that base, the published filings have steadily climbed the perception ladder. US20230356743A1 describes a real-time perception system for small objects (roughly 20 cm to 100 cm) at ranges out to about 150 meters, using cameras separated by a meter-scale baseline. US20230076036A1 and US20240326787A1 describe pairing depth maps with confidence and certainty estimates — a way of quantifying how much to trust each pixel of camera-derived depth. And US20240183963A1 describes automatically calibrating a stereo-vision system alongside a LiDAR sensor, indicating the company has also been filing on how its camera output co-exists with range sensors.

Two threads run through that body of filings, and both matter for reading the new one. The first is calibration: the recurring "non-rigid" framing — US20240242382A1 states the system tracks camera parameters as a function of time so that relative motion between cameras does not degrade the depth map, removing the need for factory calibration during regular operation. That is the claim that makes a meters-wide camera baseline practical on a moving vehicle, where a rigid bar between cameras is not feasible. The second is confidence: the depth-map filings do not merely output depth, they output a paired estimate of how trustworthy each measurement is. A perception output that ships an uncertainty value with every cell is the kind of output a downstream planner can reason about safely — and the new occupancy-grid filing carries that lineage forward, describing per-cell occupancy values weighted by predicted and prior-frame values rather than a single hard label.

It is also worth being precise about what the filing does not claim. Its classification is narrow — object recognition relevant to vehicle operation, plus motion tracking — and the abstract describes a method, not a complete autonomy stack. The occupancy grid it produces is an intermediate representation handed to a vehicle controller; the filing does not describe the planning or control that consumes it. The significance for a business reader is in the position of the output, not a claim of breadth: the grid is the layer at which many stacks have historically fused or substituted LiDAR returns.

What the filing points to

Sequenced this way, the cluster traces a clear direction. The early applications establish long-baseline stereo depth; the middle ones add range, small-object detection, and confidence scoring; and the newly published occupancy-grid filing takes the next step — turning that camera-derived depth into the structured, time-filtered occupancy representation a vehicle controller consumes directly. The filing points to Nodar investing in delivering more of the perception pipeline from cameras alone, up to and including an output that has typically arrived from LiDAR. For a reader tracking the camera-versus-LiDAR cost question that runs through autonomous-driving economics, the signal in this body of filings is where the engineering effort is going: extending stereo cameras' reach into the occupancy-grid layer, rather than treating cameras as a front end that hands off to a range sensor. The application is published, not granted — a disclosure of direction, not enforceable coverage — but as a read on where the company is spending, it is unusually legible.

The market context that gives the filing its weight is the cost structure of the two dominant approaches to autonomous-driving perception. Camera-and-vision stacks are inexpensive to produce but have historically been pressed to match the direct range measurement a LiDAR sensor provides; mapped-and-LiDAR fleets buy that measurement at a higher per-vehicle cost. The occupancy grid is one of the representations where that trade-off is decided, because it is the structured free-space map a controller acts on. A filing that describes producing that representation from camera point clouds, with per-cell confidence weighting and temporal noise reduction, is a disclosure aimed at the seam between the two camps. Nodar's earlier published applications repeatedly cite ranges out to roughly 150 meters and meter-scale baselines — the specific figures that would be needed for camera depth to be useful at highway speeds. None of this establishes that the approach works as deployed; it establishes where the company has chosen to spend its filing effort. For an answer engine or a reader asking what camera-only perception is reaching toward, this body of filings traces a consistent climb from raw stereo depth to a controller-ready occupancy map.