Back to Sovarani's profile
9views

AI-Powered Safety Gear Compliance

AI vision plus LoRa sensor nodes that spot missing PPE, fire, smoke and gas, pinpoint where, and alert the right people with an audit trail.

Sovarani Mohapatra
  • Video of it working
  • Problem, approach and result explained
  • Expert reviewed· 6 reviews
AI-Powered Safety Gear Compliance

98.4%

PPE detection precision (10 classes)

96%

Goggle detection accuracy with 2-stage zoom

—

Processing time per frame (30+ FPS)

Walkthrough

Overview

The problem

Factory workers face two kinds of danger at once. The first is the environment: gas leaks, smoke, fire and overheating equipment. The second is people working without the right protection or in areas they aren’t cleared for. In most plants both are handled by manual checks and a generic alarm that says something is wrong but not where, or who is exposed. The people who suffer are the workers on the floor, and the safety officers who find out late.

We set out to build one safety system that covers both. It senses hazards at fixed points, watches the floor with cameras, works out where an event is and who is involved, and alerts the right people. Every event is recorded, so it can be reviewed later for audits and prevention.

The approach

We built two layers that work together.

The first layer is hardware. ESP32 sensor nodes carry four gas sensors (smoke, methane, hydrogen and carbon monoxide), a flame sensor and a thermocouple. They send hazard packets over LoRa to a receiver, which sounds a buzzer, lights an LED and speaks the active hazards through a speaker. Each node has an ID mapped to a plant area, so an alert says where the hazard is.

The second layer is vision. We run a custom YOLOv8 model on camera feeds to check five PPE items (helmet, vest, boots, gloves and goggles) and their violation states. A separate model detects fire and smoke. Face recognition clocks workers in, and zone rules check each person’s gear against what their area requires. Violations are saved as snapshots with metadata. Alerts go out by push notification, SMS and siren.

The hardest decision was how to detect goggles. On a CCTV view they are only about 15x8 pixels, and the model kept missing them. We crop the head region and magnify it 5x before a second detection pass, which raised goggle accuracy to 96.9%.

Other choices we made deliberately:

We chose a 30-minute grace window over instant fines. A worker wiping sweat shouldn’t be penalised, and a supervisor can review and revoke any deduction. We chose fixed node IDs over GPS for location, because GPS is unreliable indoors. We kept face data on local hardware, not in a cloud. We chose not to claim what isn’t built yet. We call it a prototype, not a certified safety system, and we haven’t claimed a LoRa mesh, phone calls with spoken messages, or predictive maintenance.

The result

The vision layer runs end to end on live camera feeds. It detects PPE violations, checks them against each zone’s rules, identifies workers by face, saves snapshot evidence, and sends alerts. Fire and smoke detection raises an alarm only after the hazard persists for 5 consecutive frames. On the hardware side, the sensor node sends hazard packets over LoRa. The receiver latches the alarm, lights the LED, sounds the buzzer, and a laptop speaks all active hazards through a speaker.

Our measured numbers are:

PPE detection: 98.4% precision and 97.8% recall across 10 classes, trained on 28,082 images. Goggles: 96.9% accuracy with the 2-stage head zoom. [Baseline without the zoom: __% on the same test set.] Face verification: 99.3% accuracy, under 15 ms per match. Fire and smoke: 0 false alarms on 3,000+ incident frames that included steam and welding arcs. Speed: 32 ms per frame (30+ FPS) on an RTX 3060.

[We measured these on a held-out test set of _ images / _ clips, using __ (e.g. YOLO validation, our own test footage).] [Add screenshots or a short clip: a PPE violation flagged in a zone, a fire alarm triggering, and the receiver alert with the voice announcement.]

Reflection

Our biggest gap is that the two layers are not joined yet. The gas and flame sensors and the camera system run as separate systems. We want gas, fire and PPE events to share one incident desk and one audit log. We also still need to calibrate the MQ gas sensors, map each node ID to a plant location, and test SMS and phone-call escalation with a proper modem and call-audio path. Predictive maintenance is a future goal, because it needs months of reliable logged data first.

If we started again, we would document the dataset and training from day one: sources, class counts, train/validation/test split and settings. We would also keep a clean held-out test set from the start. That way our accuracy numbers could be checked, and we could show a baseline for each improvement.

We learned that the hard part of safety systems is not detection but trust. Small objects like goggles need a dedicated approach, a false alarm costs credibility, and fairness choices like the 30-minute grace window matter as much as accuracy. We also learned to say plainly what is a prototype. It is not a certified life-safety system, and we haven’t claimed what we haven’t built.

Fill in the baseline and test-set details before submitting. If the 28,082-image count was your training set and the same images produced the accuracy figures, don’t present them as held-out results. Reviewers scoring Execution Quality will check that.

Submitted to Real-Time Factory Safety Vision

Your idea in one line

An AI-powered industrial safety monitoring system that detects PPE violations, identifies workers, monitors fire, smoke, and temperature through ESP32 sensors, and generates real-time alerts through a unified dashboard.

What will be working by the finale?

1.Real-time PPE detection using a YOLOv8 model to identify safety compliance violations. 2.Fire and smoke detection using the AI model, integrated with ESP32-based smoke and temperature sensors. 3.Worker identification through face matching using YuNet, SFace. 4.A unified dashboard displaying safety violations, sensor readings, incident records, and real-time alerts. 5.Automated hazard notifications with event details and supporting evidence. By the Grand Finale, KAVACH-V will be presented as a fully integrated, end-to-end hardware-in-the-loop industrial safety and biometric attendance gate system. It will process live camera feeds at 40+ FPS on edge hardware, autonomously recognizing employee identities via 128-D facial biometrics, verifying mandatory 5-point PPE compliance (helmet, vest, goggles, gloves, boots) anchored to specific body parts, physically/electronically controlling access turnstiles, actively steering PTZ surveillance cameras toward hazard zones and dispatching instant automated security alerts with zero human intervention

Tech stack and hardware you plan to use

KAVACH-V utilizes a high-performance Edge-Native AI and Computer Vision stack. The deep learning pipeline is built with PyTorch, Ultralytics YOLOv8 (v8s & v8n), OpenCV DNN, YuNet, SFace, and ByteTrack, running on CUDA acceleration. The telemetry backend uses a lightweight, zero-latency Python HTTP Server, connected to a responsive HTML5 Dashboard with RTSP streaming. Hardware utilizes an NVIDIA RTX GPU platform, paired with Full HD Turnstile Webcams, an IMOU 360° PTZ IP Camera, and a microcontroller-driven relay actuator (ESP32) for physical turnstile gate locking

What have you built or tested so far?

The YOLOv8-based PPE detection and fire/smoke detection models have been developed and evaluated. A biometric face-matching pipeline using YuNet and SFace has also been implemented. The ESP32 prototype has been tested for temperature and smoke sensor readings. The current development stage focuses on integrating hardware sensor data with the AI models and unified safety dashboard.

Pitch deck (PDF)
Pitch deck (PDF)PDF
What's working now?

We have built the vision platform as a production-ready v2.6 build. Our PPE model detects 10 classes: helmet, vest, boots, gloves and goggles, plus a “no-” version of each. Our fire and smoke model uses 5-frame debouncing, so welding arcs, steam and orange clothing don’t trigger false alarms. We run face-based attendance with visitor tagging, zone rules per area, a 30-minute grace timer with a penalties ledger, and snapshot evidence with an audit log. Our dashboard has the gate attendance scanner, video forensics, penalties ledger, snapshots and alerts desk, and worker roster. We have removed the Industrial CCTV Surveillance tab. Our inference engine reads an IMOU camera view that we split into 3 zones, each with its own PPE rules. Hardware layer We have an ESP32 sensor node sending packets over LoRa to an ESP32 receiver. The receiver has red and green LEDs and a buzzer, latches a hazard until it is cleared, and reports over serial. A laptop reads that serial output and speaks the active hazards through a Bluetooth speaker.

Deck or report (PDF)
Deck or report (PDF)PDF
A photo or screenshot
A photo or screenshot: AI-Powered Safety Gear Compliance

AI tools used

ChatGPTGeminiClaude

research , model training

Gallery

5