constructelligence/buildvision-site-hazards

BuildVision site-hazard detector v0.1

A small (YOLOv8n, 12.2 MB ONNX) object detector for construction-site photos and video that finds workers, hard hats and hi-vis vests — worn and missing — and heavy plant (excavators, wheel loaders, dump trucks) in one pass, sized to run in a browser tab. It powers the Job Site mode of BuildVision, where missing-PPE boxes are matched to the worker they belong to and every detected machine gets an operating-radius check.

11 output classes: person, hardhat, no-hardhat, safety vest, no-safety vest, no-mask, gloves, safety shoes, excavator, wheel loader, dump truck. Of these, no-mask, gloves, safety shoes are not detected in practice — see Limitations.

Not a safety system. This is a prompt to look, not a finding. It is not safety-rated, not a substitute for a competent person's inspection, and never a compliance record.

Model output at confidence ≥ 0.35 on an image from the PPE detection 1 test split (CC BY 4.0), not seen in training.

Results

Held-out test set: the untouched test splits of 4 public construction-safety datasets, 573 images, with near-duplicates of any training or validation image removed. 3 of those sources were not in this model's training data, so their rows measure how it generalises to new sites and camera styles.

mAP50 0.442 · mAP50-95 0.241 · precision 0.714 · recall 0.432

Per class (test)

Class	Boxes	Precision	Recall	mAP50	mAP50-95
person	713	0.882	0.732	0.807	0.388
hardhat	547	0.776	0.751	0.795	0.367
no-hardhat	126	0.117	0.191	0.047	0.011
safety vest	382	0.794	0.592	0.708	0.365
no-safety vest	222	0.621	0.266	0.368	0.135
no-mask — not detected	2	1.000	0.000	0.000	0.000
gloves — not detected	248	1.000	0.000	0.000	0.000
safety shoes — not detected	268	1.000	0.000	0.000	0.000
excavator	133	0.550	0.707	0.698	0.366
wheel loader	46	0.497	0.870	0.777	0.540
dump truck	78	0.619	0.646	0.666	0.477

Per source (test)

Source	Images	mAP50	mAP50-95
Construction Site Safety	34	0.503	0.293
PPE detection 1 (not in training data)	101	0.641	0.315
PPE_Dectection v4 (not in training data)	254	0.375	0.171
excavators-czvg9 (RF100) (not in training data)	184	0.677	0.438

These are numbers on this test set. Other open PPE models report on their own datasets, so a side-by-side number is not a head-to-head comparison. Two sources label people incompletely; in their test images, people a stock COCO YOLOv8s finds (confidence ≥ 0.5) were added as person boxes so correct detections are not scored as false positives.

Use

Ultralytics (PyTorch):

from ultralytics import YOLO
model = YOLO("buildvision-hazards-v0.1.pt")
for box in model("site.jpg", imgsz=448, conf=0.35)[0].boxes:
    print(model.names[int(box.cls)], float(box.conf), box.xyxy.tolist())

onnxruntime only (no torch): pip install onnxruntime numpy pillow, then python predict_onnx.py site.jpg --out boxes.jpg. Input is 1×3×448×448 RGB in [0, 1], letterboxed with grey (114) padding; output is 1×15×anchors (cx, cy, w, h, then one score per class), NMS not included.

Browser: the same ONNX file runs in onnxruntime-web (single-threaded WASM).

Training

• Architecture: YOLOv8n from COCO weights. Fine-tuned in two stages: 24 epochs at 512 px on Construction Site Safety alone, then 6 epochs at 448 px on both datasets.
• Data: Construction Site Safety, construction-safety-gsnvb (RF100) (both CC BY 4.0): 1,304 training and 266 validation images.
• Classes dropped from the source labels: classes with one to three boxes in the whole corpus (barricade, dumpster, mask, mini-van, truck, safety net) cannot be learned and were removed before training.

Limitations

• Not detected in practice: too few training examples to learn —
  • no-mask: 36 training boxes
  • gloves: 11 training boxes
  • safety shoes: 9 training boxes These output channels exist in the model but should be ignored.
• Missing-PPE classes are the weak spot: no-hardhat and no-safety vest score well below their worn counterparts, especially on sites and camera styles not in the training data. Expect missed violations and false alarms.
• No fall protection: none of the training data labels harnesses, lanyards or edge protection.
• Small model: YOLOv8n trades accuracy for running in a browser.
• Plant operating radius is not in the model: BuildVision's radius rule is a heuristic on the detected box.

Licence

• Weights: AGPL-3.0. They were trained with Ultralytics YOLOv8, whose models and derivatives are AGPL-3.0 unless covered by an Ultralytics Enterprise licence. Using them in a network service means offering that service's source.
• Data: CC BY 4.0, attribution below.

Attribution

• Construction Site Safety — Roboflow Universe Projects, CC BY 4.0 — https://universe.roboflow.com/roboflow-universe-projects/construction-site-safety
• construction-safety-gsnvb (RF100) — Roboflow 100, CC BY 4.0 — https://universe.roboflow.com/roboflow-100/construction-safety-gsnvb
• PPE detection 1 — vincentspace, CC BY 4.0 — https://universe.roboflow.com/vincentspace/ppe-detection-1-cniwr (evaluation only)
• PPE_Dectection v4 — himanshu-bharati, CC BY 4.0 — https://universe.roboflow.com/himanshu-bharati/ppe_dectection-dtt4q (evaluation only)
• excavators-czvg9 (RF100) — Roboflow 100, CC BY 4.0 — https://universe.roboflow.com/roboflow-100/excavators-czvg9 (evaluation only)

Person pseudo-labels in the test set from Ultralytics YOLOv8s (COCO).