mancalazure/construction-ppe-safety

Construction Site PPE Detector (YOLOv8n)

Detects hard hats, hi-vis vests, masks, people, safety cones and site equipment for a worker-safety demo running on an NVIDIA Jetson Orin Nano under ZEDEDA EVE-K. A rule engine turns detections into PPE-compliance alerts and an on-device language model writes the shift summary; nothing but metadata leaves the node.

• Base: yolov8n, fine-tuned on the Roboflow Universe construction-site-safety v28 dataset.
• Classes (index order): 0 Hardhat, 1 Mask, 2 NO-Hardhat, 3 NO-Mask, 4 NO-Safety Vest, 5 Person, 6 Safety Cone, 7 Safety Vest, 8 machinery, 9 vehicle.
• Input: 640x640 (stretch). ONNX (opset 12) served on Triton's onnxruntime GPU backend.

TWO exports of the same weights -- pick the right one

This repo ships both, and they are not interchangeable. Loading the wrong one produces no error, just garbage detections.

file	input -> output	opset	exporter	consumer
model-deepstream.onnx	input [batch,3,640,640] -> [1, 8400, 6]	18	DeepStream-Yolo export_yoloV8.py	DeepStream nvinfer + NvDsInferParseYolo
best.onnx	images -> output0 [1, 14, 8400]	12	stock ultralytics (nms=False)	Triton onnxruntime, generic YOLOv8 tooling

[1, 14, 8400] is 4 box coords + 10 class scores. NMS is in neither graph -- the consumer must run it.

Files

• model-deepstream.onnx -- ONNX for DeepStream / nvinfer
• best.onnx -- ONNX for Triton / generic YOLOv8 tooling
• best.pt -- source PyTorch weights (ultralytics), for re-export
• config.pbtxt -- Triton model config (onnxruntime, GPU). Applies to best.onnx only; it is meaningless for the DeepStream graph.
• labels.txt -- class names in index order, one per line
• data.yaml -- the same class list plus dataset provenance
• logo.png -- card image

data.yaml names is authoritative for class order. Any consumer that keeps its own copy of the class list must agree exactly, including case and the space in NO-Safety Vest. Getting it wrong raises no error -- it silently mislabels every detection.

Honest limits

No validation metrics are published here, because none were measured against a held-out set for this deployment -- quoting the upstream project's numbers for this export would be misleading. What was observed in use, on one demo clip:

• Roughly one false positive per six people when "no hard hat" is inferred from the absence of a Hardhat box over a detected Person, usually from a briefly occluded hat or someone far back in frame. Raising the consecutive- frame confirmation count trades latency for quiet.
• The NO-Hardhat class itself is weak. It was trained on bare heads, so it misses caps, beanies and hoods. Inferring absence from the Hardhat class is more reliable in practice, and is what the demo does.
• No forklift and no pallet class. Heavy equipment arrives as machinery or vehicle. Say "equipment", not "forklift".
• PPE-to-person association is by box containment, so crowded frames will mis-assign a hat or vest to the wrong worker.
• "No hard hat" only fires when hats are detected somewhere in frame. Absence of detection is not evidence of non-compliance, so with a person-only model these rules go inert rather than wrong.

Validate against your own footage before quoting any accuracy figure.

Licence

Weights are CC BY 4.0, inherited from the Roboflow Universe construction-site-safety v28 dataset and the snehilsanyal/Construction-Site-Safety-PPE-Detection project -- attribution required.

Separately, the export toolchain is the YOLOv8 / ultralytics lineage, which is AGPL-3.0. That affects code, not these weights, but get it cleared before this appears in anything productised.

Usage

from ultralytics import YOLO
model = YOLO("best.pt")
model.predict("image.jpg")

Regenerating

Triton layout (this best.onnx):

docker run --rm --platform linux/arm64 -v "$PWD":/work -w /work \
  ultralytics/ultralytics:latest-arm64 \
  yolo export model=best.pt format=onnx opset=12 imgsz=640 \
              nms=False dynamic=False simplify=False

simplify=False matters: simplification pulls in onnxslim, which is not in the image, and the export dies with ModuleNotFoundError.

DeepStream layout (for nvinfer, NOT this file):

docker run --rm --platform linux/arm64 -v "$PWD":/work -w /work \
  ultralytics/ultralytics:latest-arm64 bash -c '
    pip install -q onnx onnxscript onnxslim
    git clone --depth 1 https://github.com/marcoslucianops/DeepStream-Yolo /tmp/dsy
    python3 /tmp/dsy/utils/export_yoloV8.py -w best.pt --dynamic
    python3 -c "import onnx; onnx.save(onnx.load(\"best.onnx\"), \"model.onnx\", save_as_external_data=False)"
  '

Recent torch writes split external-data ONNX -- a small .onnx plus a large .onnx.data. The onnx.save line merges them; a bare .onnx without its sidecar is a broken model.