💻 computer science

Low-Cost Stereo Vision for Robust 3D Positioning of Thin Radiata Pine Branches in Autonomous Drone Pruning

This paper proposes a low-cost, drone-mounted stereo vision system that combines YOLO-based branch segmentation with robust triangulation and outlier rejection to enable accurate 3D positioning of thin radiata pine branches for autonomous pruning, eliminating the need for expensive LiDAR sensors.

Original authors: Yida Lin, Bing Xue, Mengjie Zhang, Sam Schofield, Richard Green

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yida Lin, Bing Xue, Mengjie Zhang, Sam Schofield, Richard Green

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a drone pilot trying to trim a tree, but instead of using a human hand, you are using a robotic arm. The problem? The branches you need to cut are incredibly thin—about as thick as a pencil (10 mm)—and they are surrounded by a messy, leafy background.

For a long time, robots trying to do this job needed expensive, heavy "super-sensors" like LiDAR (which shoots laser beams to measure distance) to see these thin branches. This paper asks a simple question: Can a cheap, standard 3D camera (stereo vision) do the job just as well?

Here is how the researchers solved the puzzle, using simple analogies:

1. The "Eyes" of the Drone

The team used a single, low-cost stereo camera (like two eyes working together) mounted on a drone. Just like our eyes, this camera looks at the world from two slightly different angles to figure out how far away things are.

2. Finding the Branch (The "Spotter")

First, the drone needs to know where the branch is in the picture. The researchers tested two types of "spotting" software:

The "Heavy Lifter" (Mask R-CNN): This is a very detailed, slow, and complex system. Think of it like a master painter who takes their time to draw every single leaf.
The "Speedster" (YOLO): This is a fast, modern system. Think of it like a quick sketch artist who can spot a branch instantly.

The Surprise: On thick branches, the "Heavy Lifter" is great. But on these thin, pencil-like branches, the "Heavy Lifter" got confused and drew messy, blurry outlines. The "Speedster" (YOLO), however, was incredibly precise, drawing a tight, accurate outline around the thin branch every time. It turns out, for thin objects, the fast sketch artist is actually better than the slow painter.

3. Measuring the Distance (The "Ruler")

Once the drone knows where the branch is, it needs to know how far away it is to cut it. This is the hardest part because thin branches have very little texture (they are smooth), which confuses standard distance-measuring tools.

The researchers tested two ways to measure distance:

The "Old School" Method (SGBM): This is a traditional math formula. It works well on textured things (like a brick wall) but struggles with smooth, thin things, often creating "static" or "noise" in the distance reading.
The "AI" Method (NeRF-Supervised Deep Stereo): This is a smart neural network trained on millions of images. It is much better at guessing the distance of smooth, thin objects, producing a much cleaner "ruler" reading.

4. The "Smart Filter" (The Secret Sauce)

Even with the best AI camera, the distance reading can still be a little noisy. If you just took the average of all the pixels on the branch, a few "bad" pixels (glitches) would throw off the whole measurement, like a few bad apples ruining a basket.

The researchers invented a clever three-step filter:

Focus on the Center: Instead of looking at the whole branch (including the fuzzy edges), the system only looks at the very center line of the branch. It's like measuring a rope by looking only at its core, ignoring the frayed ends.
Ignore the Outliers: They used a mathematical trick (called MAD filtering) to spot and throw away any distance readings that were wildly wrong. It's like a bouncer at a club who kicks out the one person who is clearly not in the right line.
The Final Answer: After filtering out the noise, the system averages the remaining "good" readings to get one single, rock-solid distance number.

The Result

By combining the Speedster camera (to find the branch), the AI Ruler (to measure it), and the Smart Filter (to clean up the data), the drone can now accurately locate and measure thin branches from 1 to 2 meters away.

The Bottom Line:
The paper proves that you don't need expensive, heavy lasers to prune thin tree branches. A cheap camera, paired with smart software that knows how to ignore "noise" and focus on the center of the branch, is enough to do the job safely and accurately. This could make automated tree pruning much cheaper and safer for the forestry industry.

Technical Summary: Low-Cost Stereo Vision for Robust 3D Positioning of Thin Radiata Pine Branches

Problem Statement
Manual pruning of Pinus radiata (radiata pine) is hazardous and labor-intensive, creating a demand for autonomous solutions. Existing autonomous pruning platforms typically rely on expensive LiDAR sensors and are limited to detecting thick branches (20–40 mm), failing to address the fine, high-canopy branches (as thin as 10 mm) critical for producing clear, knot-free wood. The primary bottleneck is not the cutting mechanism, but the perception pipeline's ability to accurately detect and determine the 3D position of these thin, sparsely textured structures using low-cost hardware. This paper investigates whether a single low-cost stereo camera, without auxiliary depth sensors, can provide sufficient accuracy for this task.

Methodology
The proposed system is an integrated perception pipeline comprising two parallel stages: branch segmentation and depth estimation, fused via a novel triangulation algorithm.

Branch Segmentation:
The study evaluates instance segmentation models to generate pixel-level masks of branches. Two families are compared:
- Two-stage detectors: Seven variants of Mask R-CNN (using ResNet-50/101 and ResNeXt-101 backbones with C4, DC5, and FPN configurations).
- One-stage real-time detectors: Variants of YOLOv8 (n, s, m, l, x) and YOLOv9 (c, e).
  The models are trained on a custom dataset of 71 stereo image pairs captured indoors with a ZED Mini camera.
Depth Estimation:
The pipeline compares traditional and deep-learning-based stereo matching methods to generate disparity maps:
- Traditional Baseline: Semi-Global Block Matching (SGBM) with Weighted Least Squares (WLS) post-filtering.
- Deep Learning: Six architectures including cost-volume networks (PSMNet, ACVNet, GWCNet, MobileStereoNet), iterative refinement networks (RAFT-Stereo), and NeRF-Supervised Deep Stereo.
- Cross-Dataset Analysis: Experiments include fine-tuning PSMNet on urban driving datasets (KITTI 2012, KITTI 2015) and Scene Flow to quantify the domain gap between urban benchmarks and forestry scenes. Monocular depth models (MiDaS, Depth Anything) are evaluated only as a sanity check for metric scale ambiguity.
Fusion and 3D Positioning (Main Contribution):
To convert a segmentation mask and a noisy disparity map into a single robust distance, the paper introduces a centroid-based triangulation algorithm with Median-Absolute-Deviation (MAD) outlier rejection:
- Centroid Sampling: Instead of averaging all pixels in the mask (which includes noisy silhouette/background pixels), boundary points are grouped into triangles. The centroids of these triangles serve as sampling locations, biasing samples toward the branch's medial line where matching is more reliable.
- Neighborhood Expansion: Each centroid is expanded by neighboring sample points to reduce variance.
- MAD Filtering: Outliers are rejected based on the Median Absolute Deviation of the depth values within the sample pool. Only values within $median \pm k \cdot MAD$ (where $k=3$ ) are retained.
- Final Estimation: The final branch-to-camera distance is the mean of the retained, filtered depth values.

Key Results

Segmentation Performance: On the branch dataset, YOLOv8 and YOLOv9 variants significantly outperformed Mask R-CNN in mask quality. While Mask R-CNN achieved high bounding box accuracy (77–88 mAPbox), its mask accuracy collapsed below 12 mAPmask. In contrast, YOLO variants achieved 98.7–99.6 mAPbox and 77.1–82.0 mAPmask. The authors attribute this to YOLO's high-resolution prototype prediction versus Mask R-CNN's fixed low-resolution mask upsampling, which struggles with thin structures.
Depth Estimation:
- Domain Gap: Fine-tuning PSMNet on urban driving datasets (KITTI) degraded performance on thin branches, confirming a significant domain gap.
- Architecture Comparison: NeRF-Supervised Deep Stereo produced the cleanest disparity maps for branches without requiring in-domain ground truth, outperforming SGBM and other deep networks in preserving branch edges.
- Fusion Efficacy: The centroid-based triangulation with MAD filtering successfully converted noisy disparity maps into stable distance estimates. Qualitative histograms showed that the NeRF-Stereo pipeline produced depth estimates more concentrated around the true distance (1m, 1.5m, 2m) compared to SGBM.
Ablation: Removing MAD filtering or using full-mask sampling (instead of centroid sampling) resulted in significant distance drift (tens of centimeters) due to background bleed and silhouette artifacts.

Significance and Claims
The paper claims that low-cost passive stereo vision, when coupled with prototype-based one-stage segmentors and the proposed fusion algorithm, is a viable perception front-end for autonomous pruning of fine branches (down to 10 mm).

Novelty: The primary contribution is the coupling of stereo segmentation with a centroid-based triangulation and MAD outlier rejection, which addresses the specific challenges of sparse texture, thin structures, and noisy disparity values in forest scenes.
Feasibility: The study demonstrates that auxiliary depth sensors (like LiDAR) are not strictly necessary for branch positioning at working distances of 1–2 m, provided the perception pipeline is robust to noise.
Limitations and Future Work: The authors modestly note that the current dataset is limited to indoor conditions and lacks wind motion or dense canopy clutter. They identify the need for outdoor field collection with LiDAR reference data to provide quantitative end-to-end accuracy figures and the development of multi-task architectures to reduce on-board latency for real-time drone operation.