INTELSY PPE Detection

Checking that everyone on a construction site or factory floor is wearing their protective equipment is usually done by eye, and it is easy to miss someone. The project explores automating that check with object detection, as decision support for a safety officer rather than automatic enforcement.

• Detects 14 classes, including Hardhat, Safety Vest, Mask, Gloves and Goggles and their NO- counterparts, plus Person, Ladder, Safety Cone and falls.
• Runs on single images, folders of images and live webcam video, drawing bounding boxes with class labels and confidence.
• Classes can be kept or hidden at inference time, with adjustable confidence and IoU thresholds.
• Each run writes a plain-language summary of what was detected and whether required PPE is missing.
• A ready checkpoint is included, so the webcam demo runs without training.
• One-command reproducible training run, with a script that builds the ablation table.

Measured against a baseline, not just reported

A non-deep-learning baseline that predicts the most common class for every box scores 0.41 accuracy and 0.04 macro-F1, which gives the detector's numbers something to be compared against. Short ablation runs on image size and data fraction are tabulated alongside.

The final model was trained in Colab

Training continued from an earlier checkpoint for 30 epochs at 640px on Google Colab. The candidate checkpoint reached 0.79 mAP@0.5 and 0.52 mAP@0.5:0.95 on the test split, while short CPU runs stayed for quick reproducibility.

Decision support, not enforcement

The release notes are explicit that detections vary with occlusion, motion blur and small objects, so output is framed as an aid for a person checking compliance.

• Python/
• YOLOv8 (Ultralytics)/
• PyTorch/
• OpenCV/
• Roboflow/
• Jupyter/
• Google Colab/
• pandas & scikit-learn