Computer-vision pipeline for non-contact measurement of utility pole geometry: diameter from walking video and tilt angle from field images.
Traditional utility pole inspection requires physical contact with specialised calipers or laser rangefinders. This project replaces those methods with an AI photogrammetry pipeline operating on standard smartphone video and field photographs — no special hardware needed.
Two independent measurement capabilities:
| Capability | Input | Output |
|---|---|---|
| Pole diameter | Walking video (smartphone) | Diameter in inches & mm, per-frame annotated video |
| Pole tilt angle | Single field photograph | Tilt in degrees from vertical, annotated image + methodology figure |
Real-time diameter estimation from a smartphone walking video. Blue overlay = pole mask · Red overlay = inspection badge (1.25 in reference) Green line = measured diameter at badge height.
flowchart LR
V[Video] --> YW[YOLO-World\nPOLE + BADGE detection]
YW --> SAM[EfficientSAM\ninstance segmentation]
SAM --> DM[Pole mask slice\nat badge midpoint]
DM --> SC["Diameter = 1.32 × (pole_px / badge_px)"]
SC --> OUT[Diameter in inches & mm]
- Reference-object detection — YOLO-World detects
POLEandBADGE(a 1.25-inch inspection badge affixed before filming) simultaneously. - False-positive pruning — heuristics filter by aspect ratio, area bounds (4–75 % of frame), and spatial containment of badge within the pole box.
- Pixel-accurate segmentation — EfficientSAM masks the pole; the mask is sliced at the badge midpoint to measure the true cross-sectional width.
- Depth-calibrated scaling — a scaling factor of
1.32converts the pixel ratio to real-world inches (calibrated by linear regression on 35 images). - Aggregation — the median across all video frames is the final estimate.
| Metric | Value |
|---|---|
| Mean Absolute Error | 0.47 in (12.0 mm) |
| RMSE | 0.57 in (14.6 mm) |
| Best single estimate | 0.022 in (0.56 mm) |
| Diameter range tested | 9.0 – 16.0 in |
Distribution of per-frame diameter errors across the 35-image validation set.
Linear regression of depth-scaling factor vs. badge-to-image area ratio.
# Single video — ground-truth diameter optional (enables error overlay)
uv run python src/main.py -i data/sample/sample_1.mp4 -d 9.5
# Default sample
uv run python src/main.py
4-panel methodology figure: (1) input image · (2) EfficientSAM pole mask · (3) medial-axis skeleton · (4) Hough line candidates with estimated tilt.
flowchart LR
I[Field photo] --> YW[YOLO-World\nPOLE detection]
YW --> SAM[EfficientSAM\ninstance segmentation]
SAM --> MA[Medial-axis\ntransform]
MA --> HT[Hough line\ntransform]
HT --> OUT["Tilt angle\n(degrees from vertical)"]
- Detection & segmentation — same YOLO-World + EfficientSAM stack as diameter estimation.
- Medial-axis transform —
skimage.morphology.medial_axisthins the binary pole mask to a single-pixel centreline skeleton. - Hough line transform —
skimage.transform.hough_linefinds dominant straight-line orientations. The median of the top-5 peaks is the pole axis. - Angle convention — Hough
thetaequals deviation from vertical:0°= perfectly upright;+= leans right;−= leans left.
| File | Content |
|---|---|
<name>_tilt.jpg |
Clean annotated image with angle arc and degree label |
<name>_analysis.jpg |
4-panel methodology figure |
# Single image
uv run python src/tilt.py -i data/sample/pole_tilt/leaning_pole_01.jpg \
-o result/pole_tilt/leaning_pole_01_tilt.jpg
# Batch — process a folder, write CSV summary
uv run python - <<'EOF'
import sys; sys.path.insert(0, "src")
from AI import AI
from tilt import run_tilt_batch
ai = AI()
run_tilt_batch(
input_dir = "data/sample/pole_tilt/",
output_dir = "result/pole_tilt/",
ai = ai,
results_csv = "result/pole_tilt/tilt_summary.csv",
)
EOF├── src/
│ ├── AI.py # YOLO-World + EfficientSAM model wrapper
│ ├── diameter.py # Width/diameter estimation class
│ ├── tilt.py # Tilt angle estimation + batch runner
│ ├── video.py # Frame extraction / video reconstruction (ffmpeg)
│ └── main.py # CLI entry points (pole_diameter, pole_tilt)
│
├── data/sample/
│ ├── sample_1–3.mp4 # Width estimation input videos
│ ├── manual_diameter.csv
│ ├── pole_tilt/ # 22 leaning-pole field images (leaning_pole_NN.jpg)
│ └── pole_width/images/ # 7 still images for diameter testing
│
├── results/
│ ├── width_estimation/
│ │ ├── pole_width_result_01–02.mp4 # annotated output videos
│ │ └── analysis/ # error CSV + 3 analysis plots
│ └── tilt_estimation/ # sample tilt methodology figures
│
├── assets/
│ ├── pole_width_demo.gif # animated demo for README
│ ├── pole_width_annotated.jpg # sample annotated frame
│ └── pole_tilt_analysis.jpg # sample 4-panel tilt figure
│
├── models/README.md # model download instructions
├── pyproject.toml # uv / PEP 517 project spec
└── requirements.txt # pip-compatible dependency list
- Python 3.11+
- uv —
curl -LsSf https://astral.sh/uv/install.sh | sh - ffmpeg —
brew install ffmpeg(macOS)
git clone https://github.com/ashish-code/utility-pole-ai-photogrammetry.git
cd utility-pole-ai-photogrammetry
# Install all dependencies into an isolated venv
uv sync
# Download model weights (see models/README.md for details)
wget "https://huggingface.co/spaces/SkalskiP/YOLO-World/resolve/main/efficient_sam_s_cpu.jit"
uv run python -c "from ultralytics import YOLOWorld; YOLOWorld('yolov8x-world.pt')"AI.py selects CUDA automatically when available.
| Device | Frame throughput |
|---|---|
| NVIDIA RTX 3090 | ~3 frames/s |
| Apple M-series (CPU) | ~0.2 frames/s |
| Library | Role |
|---|---|
ultralytics |
YOLO-World open-vocabulary detection |
torch / torchvision |
EfficientSAM inference |
supervision |
Detection parsing & annotation helpers |
scikit-image |
Medial-axis transform, Hough line transform |
opencv-python |
Image I/O, drawing, ffmpeg video ops |
matplotlib |
Methodology figures |
pandas |
Batch results CSV |
Developed in collaboration with Osmose Utilities Services to explore non-contact AI approaches to field inspection. The badge-based photogrammetry method provides a scalable, low-cost alternative to contact measurement for utility pole condition surveys.
MIT — see LICENSE.