Back to Bibhudutta's profile
11views

NETRAX

Edge-AI safety that checks 7 PPE items per worker and spots fire on existing CCTV, fully on-site with no cloud, and alerts the supervisor's phone.

Bibhudutta Biswal
  • Video of it working
  • Problem, approach and result explained
  • Public source code
  • Answered the full brief· 13 of 13
  • Expert reviewed· 6 reviews
NETRAX

—

95.9% PPE detection accuracy (mAP50, hard test set)

—

0.28 s Camera-to-dashboard alert time

—

4 Live cameras on one laptop GPU, no cloud

Walkthrough

Overview

The problem

In industrial environments, especially in high-risk areas like welding bays, assembly lines and warehouses, ensuring worker safety is a major challenge. Workers may enter restricted areas without proper authorization or fail to wear essential protective equipment such as helmets, safety glasses, gloves and safety shoes. Identifying these violations manually through supervisors and CCTV monitoring can be difficult, time-consuming and prone to human error. Another major concern is the delayed detection of fire and smoke. If a hazardous situation is not identified and communicated immediately, it can put workers' lives and factory operations at serious risk. Supervisors also need to monitor worker attendance, zone access, PPE compliance and safety conditions across different areas, often using separate systems. The question we wanted to address was: How can we make industrial workplaces safer by identifying workers, verifying their PPE and zone authorization, and detecting fire hazards in real time, while keeping supervisors informed through a single monitoring platform? Our goal is to develop an integrated system that reduces dependence on manual supervision, identifies safety violations quickly and enables faster action during emergencies.

The approach

NETRAX is an edge-AI safety system that checks every worker’s PPE and watches for fire and smoke. Our biggest strength is that it runs fully on-site on one laptop GPU, with no cloud. Video never leaves the site, and alerts are queued and still go out when the internet returns. It works with existing CCTV cameras over RTSP, as well as webcams and phones. Data: public PPE and fire/smoke datasets, plus our own camera frames that we labelled by hand, including hard cases like hair, flash and phone screens. Models: YOLOv8-pose finds people and their body points. Our own NETRAX-YOLO, with an extra small-object layer, detects 7 PPE items. There is also a fire/smoke model, a shoe classifier and face recognition for names. A simple detector only finds “a helmet somewhere”. We match each item to the right body part of the right person, and mark it worn, missing or “can’t see”. An alert needs 6 of 10 frames, and fire is confirmed by our sensor station.

The result

We measured mAP50 on a held-out hard test set the models never saw:

PPE: our final model scored 0.959 (precision 0.926, recall 0.936), against 0.940 for the previous version. Fire/smoke: 0.870, against 0.161 for the previous version. Our first fire model scored 0.982 on an easy test but only 0.706 on validation,. Shoe classifier: 14 of 15 correct. We also uploaded our own datasets. This is what helped our model to train accordingly and detect minor abnormalities.

Live, measured from our logs: about 70 ms per frame, 4 cameras running together, 0.28 s from camera to dashboard, over 1,700 Telegram alerts and over 900 e-mails delivered, 130 self-test checks passing.

It still gets these wrong:

need a more powerful GPU for multiple camera detection. one violation sometimes created duplicate alerts, until acknowledged by a supervisor. phone cameras on busy Wi-Fi can drop to 1 fps.

Reflection

We would design the dashboard for supervisors first, since ours became crowded, and test on real CCTV much earlier. But after our Round 2 evaluation we rectified it. We would try to categories more zones specifying each factory need. We would work with factories and industry on successfully deploytion with pre exciting CCTv. What we learned:

A high score on an easy test can lie; real hard test sets matter. False alarms destroy trust, so “can’t see” is better than guessing. Edge AI solves both privacy and network delay. Reliability work like retries, offline queues and self-tests matters as much as the model.

Submitted to Real-Time Factory Safety Vision

Your idea in one line

NETRAX is an AI-powered safety system that monitors workers’ safety gear, detects fire and smoke in real time, and ensures only authorized workers enter high-risk zones. It identifies the most critical danger first and delivers instant alerts and voice warnings by name, all running on-site without relying on the cloud. Using ESP32 controllers, environmental sensors, RFID readers, CCTV cameras, and alarm hardware, it identifies the most critical danger first and delivers instant alerts and voice warnings by name, all running on-site without relying on the cloud.

What will be working by the finale?

Future Scope of NETRAX Enhanced PPE Detection: Extend AI detection to identify additional safety gear such as harnesses, earmuffs, welding masks, and respirators, while verifying correct usage and detecting damaged or counterfeit equipment. UWB-Based Worker Tracking: Integrate Ultra-Wideband (UWB) tags for precise real-time worker location tracking in crowded industrial zones and time-based access control for restricted areas. Improved CCTV Monitoring: Enhance the AI model for night-time and low-light conditions, crowded environments, and far-distance or overhead camera views, with continuous improvement through supervisor feedback. Smart Wearable Safety: Develop compact sensor modules integrated into helmets and safety equipment, reducing loose wiring, and explore wearable cameras for first-person safety monitoring through the dashboard. Measurable Safety Compliance: Introduce detailed compliance analytics and reporting to track PPE usage, identify recurring safety violations, and support data-driven industrial safety decisions.

Tech stack and hardware you plan to use

## Tech Stack - **AI/Computer Vision:** Python 3.11, PyTorch, YOLOv8 (PPE, fire/smoke, pose), ByteTrack, OpenCV (YuNet/SFace), ONNX. - **Training & Evaluation:** Google Colab, Kaggle, Roboflow, D-Fire, custom videos; Precision, Recall and mAP. - **Backend & Database:** FastAPI, Uvicorn, SQLite, WebSockets, CSV/PDF reports, offline alert buffering. - **Dashboard & Alerts:** HTML, CSS, JavaScript, live CCTV, analytics, heatmaps, voice alerts, siren and Telegram notifications. - **Firmware:** Arduino C++ on ESP32; Wi-Fi, HTTPClient, MFRC522 and sensor libraries. ## Hardware - **Edge AI Unit:** Laptop with NVIDIA GTX 1650 (4 GB GPU). - **Cameras:** Laptop webcam and two Android phones as IP cameras. - **Safety Station:** ESP32, DHT22, MQ-2, MQ-135, flame sensor, PMS5003, thermistor, FSR/touch sensors, LEDs, buzzer and siren relay. - **Access Control:** Separate ESP32, RC522 RFID reader, worker cards and buzzer. - **Network:** Shared Wi-Fi hotspot connecting cameras, laptop and ESP32 units. ## Key Integration - Camera-based fire/smoke detection is verified by hardware sensors before siren activation. - RFID and consent-based face recognition support worker identification and access alerts. - AI video processing runs locally; alerts and thumbnails are shared with the dashboard.

A diagram, sketch or mock-up of your idea
A diagram, sketch or mock-up of your idea: NETRAX
What have you built or tested so far?

NETRAX has integrated six AI models for real-time PPE and fire/smoke detection, supported by a custom YOLOv8 model and a dataset of 1,314 labelled frames from 27 videos. We have implemented face recognition, three-zone monitoring, ESP32-based environmental sensing, RFID access control, priority-based alerts, voice warnings, and supervisor dashboards. Testing achieved 0.96 mAP50 for PPE detection and 0.87 mAP50 for fire detection, rejected 31 out of 31 tested fake-smoke cases, and demonstrated approximately 70 ms inference per frame, with dashboard alerts in under 0.2 seconds. Automated logic, hardware, and access-control tests passed.

Pitch deck (PDF)
Pitch deck (PDF)PDF
GitHub repository
github.com/Cbdey/NETRAX
What's working now?

Working: PPE detection for 7 items: helmet, vest, gloves, goggles, mask, safety shoes and face cover. It checks each worker item by item and marks every item as worn, missing or “can’t see”. Fire and smoke detection on camera, cross-checked with our ESP32 sensor station (flame, gas, smoke and temperature sensors). 4 cameras running live at the same time on one laptop GPU, with about 0.28 s from camera to dashboard. Entrance check: a worker is only marked PERMITTED when all their required PPE is visible. Face recognition puts the worker’s name and ID on alerts. There are also restricted-zone alerts and RFID attendance. Alerts go to the dashboard, Telegram with a photo, e-mail, the laptop speaker by voice, and a Twilio phone call for fire. Everything runs on-site and still works without internet: messages are queued and sent when it comes back. Partly working: Real CCTV: RTSP support is ready and we are connecting an AXIS IP camera, but we haven’t tested it for long hours. Phone cameras sometimes drop to 1 fps on busy Wi-Fi. Fire calls go out, but don’t repeat yet until someone acknowledges them. E-mail sometimes failed because we hit Gmail’s daily sending limit. Didn’t get to: WhatsApp alerts. The GSM/GPS hardware module for location (the software side is ready). A long trial on a real factory site.

GitHub repository
github.com/Cbdey/NETRAX
Deck or report (PDF)
Deck or report (PDF)PDF
A photo or screenshot
A photo or screenshot: NETRAX

AI tools used

ChatGPTClaudeGitHub Copilot

CLAUDE- CODING AND LOGIC SETUP CHATGPT-DEBUGGING AND HARDWARE PART ANTIGRAVITY (CLAUDE-FANLE 5.2 )-BACKEND LOGIC DESIGN AND API INTEGRATION GITHUB COPILOT-GIT REPOSITORY INITIALISATION AND PUSH COLAB- MODEL TRAINING KAGGLE - DATASET COLLECTION AND PART MODEL TRAINING ROBOFLOW- DATASET COLLECTION

Links & files

Artifacts

1

Gallery

5