Construction Site PPE Compliance Monitoring with Privacy-First Video

The problem sites have with safety compliance

On construction sites, workshops and factory floors, helmets and high-visibility vests are required and routinely not worn, especially in heat, at the end of a shift, or when a supervisor is not looking. Safety officers walk the site a few times a day; the rest of the time nobody knows. When an accident happens, the question is always whether the rules were being followed, and there is rarely any data.

Many sites already have CCTV cameras for theft. That footage could tell a safety officer where and when compliance drops, without adding hardware. The obstacle is that watching hours of video is impossible, and recording faces for monitoring raises reasonable privacy concerns from workers.

What already exists and the gap

• Intenseye and Protex AI sell video analytics for workplace safety to large industrial sites, with enterprise pricing and deployment projects.
• viAct focuses on construction site monitoring with similar enterprise sales.
• Ultralytics publishes a Construction-PPE dataset (1,416 images, 11 classes including missing gear) and training tools, which makes detectors easy to train but stops at the model.

The gap is an affordable, privacy-first system for mid-sized contractors and factories that turns existing cameras into anonymous compliance data rather than surveillance of individuals.

What you build

A video pipeline and a dashboard. The pipeline reads RTSP streams or recorded files, runs a detector on sampled frames, tracks people across frames, associates each person with helmet and vest detections, and applies zone rules. Before any snapshot is saved, faces and heads are blurred. The dashboard shows compliance per zone and hour, and lets the safety officer review alert snapshots and mark them as correct or false alarms.

You train on public PPE datasets first, then fine-tune on frames from a site you can access with permission, such as a campus construction project. Label a few hundred frames yourself; local helmets, clothing and camera angles differ from public data.

The technically interesting part

Going from frame detections to trustworthy events. A per-frame detector will flicker: a helmet is detected, missed, detected again. Counting per frame gives absurd violation numbers. You need tracking (ByteTrack or similar), association of PPE to the right person when people overlap, and a temporal rule ("no helmet for at least three seconds inside the zone") before raising an event. Show your evaluator the difference: violation counts and precision per frame versus per track with the temporal rule, on the same annotated clips. That comparison is the heart of the project.

The privacy design is the second strong point: show that no stored artefact contains an identifiable face, and document what is and is not kept.

How to evaluate it

• Detection mAP on the public test split and on your own labelled frames.
• Event-level precision and recall on at least 30 minutes of annotated video, per frame versus per track with temporal rules.
• False alarm rate per camera hour at the chosen settings.
• Throughput: frames per second per camera on a laptop GPU and on CPU.
• A privacy check: a sample of stored snapshots reviewed by two people who try to identify individuals.
• Feedback from one safety officer or site engineer on whether the dashboard would change their rounds.

Suggested plan

1. Month 1: permissions for a local site, dataset review, baseline training on public data.
2. Month 2: collect and label local frames, fine-tune, compare model sizes.
3. Month 3: tracking, PPE association and temporal rules.
4. Month 4: blurring, event storage and zone configuration.
5. Month 5: dashboard and alert review workflow.
6. Month 6: annotated clip evaluation, safety officer feedback, write-up.

Path to a product

The first customer is a mid-sized contractor or a factory with existing CCTV and an internal safety officer. Charge per camera per month, installed on a small on-site box so video never leaves the premises, which also answers privacy objections. Scaling means more rule types (restricted zones, forklifts near people) using the same tracking backbone.

Scope warning

Do not identify individual workers or build any disciplinary feature. Do not stream video to the cloud. Do not try to detect every item of PPE; helmet and vest done well is enough. Always get written permission before recording on any site.

Learn what you need on Sefism

• Machine Learning roadmap : training, validation and fine-tuning for the detector.
• Machine Learning cheat sheet : mAP, precision and recall for detection tasks.
• System Design roadmap : designing the video pipeline and event storage.
• Google Colab : free GPUs for training YOLO models.
• EU Data Protection Rules for Business : a useful frame for data minimisation in workplace video.

Get help with this idea

The two areas where students usually need guidance are the event logic and the privacy design. A mentor can help you define both before you collect footage. Bring your draft scope, the dataset you plan to use and your evaluation table, and we can look at them together. The build is yours; the guidance is on scope, stack and how to measure it. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session