ESP32-CAM PPE Detection for Construction Sites

PPE compliance on construction sites is still enforced by supervisors walking the site with a clipboard. A supervisor cannot watch every corner continuously, violations go unrecorded, and there is no auditable evidence trail when an incident is reviewed. Fixed CCTV exists on larger sites, but nobody watches every feed. The engineering fix is to give the cameras a brain: capture frames from cheap ESP32-CAM nodes, stream them over the site Wi-Fi, and run a trained object detector that spots workers and flags missing helmets in near real time. This project builds the complete loop — ESP32-CAM firmware for frame capture and streaming, a YOLOv8s detector fine-tuned on annotated construction-site imagery for the helmet and no-helmet classes, and a web dashboard with the annotated live feed, compliance statistics, violation alerts and daily trend charts. The training notebook logs mAP, precision and recall during the build, so the report presents the student's own evaluation, not a pre-claimed number.

1. Each ESP32-CAM node captures JPEG frames at a configured interval and serves them over Wi-Fi to the site server.
2. The Flask/FastAPI server pulls frames from all registered camera nodes and batches them for inference.
3. The fine-tuned YOLOv8s model detects workers in each frame and classifies each detection as helmet or no-helmet.
4. Non-maximum suppression removes duplicate boxes; detections above the confidence threshold are drawn on the frame with labels.
5. The annotated frame, per-camera compliance rate and violation events are pushed to the dashboard over WebSockets.
6. No-helmet detections are written to the violation log with timestamp, camera ID, confidence and the evidence frame.
7. A nightly job aggregates the day's detections into compliance trend charts and per-zone summaries.

Tech stack:

• ESP32-CAM (AI-Thinker) with Arduino-core capture/streaming firmware
• Python, Ultralytics YOLOv8s (training and inference), OpenCV
• Flask/FastAPI server + WebSocket live feed
• Single-file HTML/CSS/JS dashboard (annotated feed, stats, violation log, trend charts)
• Hard Hat Workers-style annotated dataset (helmet / head / person classes)
• SQLite event store for violations and daily aggregates

• [Helmet / no-helmet detection] YOLOv8s fine-tuned on annotated site imagery locates workers and classifies helmet usage, drawing labelled boxes with confidence scores.
• [ESP32-CAM capture nodes] Camera firmware grabs frames and streams them over Wi-Fi; nodes are cheap enough to cover multiple site zones.
• [Split architecture] The ESP32-CAM captures and streams only — inference runs on the site PC/laptop, which is stated explicitly so the design is honest about hardware limits.
• [Live annotated dashboard] Real-time feed with detection overlays, per-camera compliance percentage and current shift violation count.
• [Violation log with evidence] Every no-helmet detection is logged with timestamp, camera ID, confidence and the saved frame as evidence.
• [Compliance trends] Daily and weekly compliance charts plus a per-zone heatmap-style breakdown of violation counts.
• [Configurable alerting] Confidence threshold and alert cooldown adjustable from the dashboard; optional buzzer/email alert hooks documented.
• [Complete training pipeline] Dataset preparation, augmentation, training and evaluation notebooks included; mAP, precision, recall and PR curves logged during the build.

• ESP32-CAM firmware (frame capture, Wi-Fi streaming, multi-node registration)
• YOLOv8 training and evaluation notebooks (dataset prep, augmentation, metrics logging)
• Detection server code (frame ingestion, inference, WebSocket feed, violation logging)
• Web dashboard (live annotated feed, compliance stats, violation log, trend charts)
• Wiring and deployment guide (node placement, Wi-Fi coverage, power options)
• Project report PDF (background, dataset analysis, methodology, evaluation, limitations)
• PPT presentation for final review
• Viva Q&A preparation document (YOLO, mAP, ESP32-CAM limits, streaming, evaluation)
• Setup guide (environment, dataset download, training, running the system)

• The ESP32-CAM cannot run the detector itself — inference needs a site PC or laptop; the report states this split architecture explicitly and explains why.
• Detection quality depends on camera angle, lighting and occlusion; backlit, night-time or heavily occluded workers are missed more often, and the evaluation section documents this.
• Public datasets may not match your site's helmets, vests and backgrounds; the notebook includes a fine-tuning path on custom-collected site images.
• Wi-Fi coverage limits node placement; frames drop when the link is weak, and the dashboard marks stale cameras clearly.
• This is a compliance-assist tool, not a safety-certified system — it does not replace supervision, site safety procedures or legal PPE obligations.
• Final mAP and FPS are measured on the student's own hardware during the build; the design target is a target, not a guaranteed result.