1
Correct the perspective
INVERSE PERSPECTIVE MAPPINGBIRD'S-EYE VIEW
Map image pixels onto an assumed ground plane. Rows become easier to compare in metres.
Camera height and pitch matter. Tall plants violate the flat-ground assumption. CAMERA-GUIDED AGRICULTURE
From pixels to a path.
See how a forward-facing camera becomes a crop-row estimate—and when a robot should stop trusting it.
INTERACTIVE ENGINEERING DEMO
Perspective correction. Row evidence.
A navigation decision you can inspect.
02 OPEN THE PERCEPTION PIPELINE
Every decision starts with evidence.
These views are calculated from the camera pixels.
Change a condition and watch the evidence change.
2
Keep the vegetation
COLOUR-BASED SEGMENTATIONDETECTOR INPUT
Retain green plant evidence and suppress much of the background before line detection.
Green is a colour cue. It does not distinguish crops from weeds.3
Let the rows vote
CONSTRAINED HOUGH ESTIMATIONLINE EVIDENCE
A candidate row pair is checked against its support and the expected spacing.
Turn coherent line evidence into a corridor, lateral offset and heading correction.
Weak evidence requests a hold. A straight-line model has limits on curved rows.03 TEST THE ASSUMPTIONS
Better questions. Better deployment decisions.
Compare realistic failure modes before planning a trial.
Apply changes to reset the simulation with your settings.
Field conditions
Change the evidence the camera receives.
Camera & calibration
Separate physical geometry from the assumed model.
Detector & motion
Inspect the choices behind the steering command.
YOUR IMAGE. THE SAME PIPELINE.
Bring a field image into the conversation.
Analyse a local photograph using the current calibration and row-spacing settings. Image mode is stationary analysis; it does not command the simulated robot.
Your image is processed locally in this page. For metric results, use an undistorted image with no camera roll and known camera height, pitch and vertical field of view.JPEG · PNG · other browser-supported images
TAKE THE EVIDENCE WITH YOU
Show the method. Share the result.
Save an annotated perception frame or a machine-readable record of settings and estimates.
04 ENGINEERING CONTEXT
Useful insight needs honest boundaries.
A transparent demonstration of the perception chain.
Field performance must be measured in a real deployment.
What this demonstrates
A forward RGB image is rectified to a ground-plane view. Vegetation pixels vote for candidate lines. A row-spacing prior helps select a corridor, which supplies lateral and heading errors to a simulated controller.
What can go wrong
Weeds, missing plants, shadows, curved rows and pitch errors affect the evidence. Good-looking line fits do not guarantee correct metric geometry. The support score is a heuristic, not a calibrated confidence probability.
What a deployment needs
Camera calibration, crop-specific validation and representative field recordings. Operational robots also need obstacle detection, independent stop functions and a tested navigation system. This demo is not a safety system.
Method, references & technical scope+
A practical engineering discussion
For growers, this workbench helps frame a camera trial: row spacing, visibility, calibration and conditions that warrant stopping. For researchers and decision-makers, it makes the transformation from pixels to steering evidence explicit.
The synthetic scene supports controlled comparison. It is not a benchmark of real crop images, a product performance guarantee or a substitute for field evaluation. Camera perspective correction assumes a pinhole model and flat ground; vegetation above that plane introduces parallax. Imported images also assume zero camera roll, a centred principal point, square pixels and negligible lens distortion.
Implemented method. Normalized excess green, (2G − R − B) / (R + G + B), supplies vegetation evidence. A weighted Hough transform searches row orientations within ±22°. Non-maximum suppression and robust line refinement produce candidates; both rows must be independently supported across depth bands. The default spacing prior is 1.5 m. A pure-pursuit controller turns the selected corridor into a steering command. The pipeline runs in this page’s own JavaScript; it does not require an OpenCV runtime.
The demonstration uses a constrained Hough-based method. The research referenced in the source slide uses a different approach, involving HDBSCAN and RANSAC; this page does not reproduce or validate that paper's method.
Further reading
- OpenCV · Homography and perspective geometry ↗
- OpenCV · Hough line transform ↗
- Khan et al. (2024) · Crop-row detection study ↗
Software notices
ADAPTIVE AGROTECH — ROWSIGHT 3D Third-party notices Three.js r128 and OrbitControls Source: https://github.com/mrdoob/three.js License: MIT MIT License Copyright 2010-2021 Three.js Authors Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. Original Adaptive AgroTech branding is preserved from the original supplied logo. All field and robot geometry is procedural. No remote dependencies are used.