Autonomous AI-Driven Industrial Worker Safety and PPE Compliance Monitoring System: A Final Year Project

Abstract

Occupational hazards in manufacturing plants, construction sites, and hazardous industrial environments remain a critical concern, leading to severe injuries and non-compliance fines. Traditional manual safety inspections are labor-intensive, continuous monitoring across multiple high-risk zones is difficult, and human supervisors are subject to fatigue. This project presents an autonomous, real-time Personal Protective Equipment (PPE) detection and compliance monitoring platform powered by advanced deep learning algorithms and computer vision pipelines.

By leveraging optimized object detection architectures, edge computing capabilities, and spatial association techniques, the system continuously analyzes video streams from existing IP/CCTV cameras. It detects whether workers are correctly equipped with hard hats, high-visibility vests, safety goggles, and protective gloves while simultaneously detecting restricted area intrusions. The end-to-end framework incorporates automated alert generation, statistical compliance tracking, and structured reporting to ensure full operational safety.

Project Overview

The Automated Industrial PPE Compliance Verification platform provides enterprise-grade safety auditing using visual intelligence. The pipeline ingests high-definition Multi-RTSP (Real-Time Streaming Protocol) video streams, pre-processes visual frames, and applies custom-trained convolutional neural network (CNN) and Vision Transformer models to recognize human operators and associate specific gear items to each detected worker.

In addition to simple detection, the system applies spatial bounding-box overlay logic to perform 'PPE-to-Person' association. If a worker is detected without mandatory gear inside a designated danger zone, the platform triggers an instantaneous visual and audible alert while logging the incident with timestamped visual evidence into a centralized database for OSHA safety audits.

Problem Statement

Industrial workplace accidents cost businesses billions of dollars globally and cause irreversible physical harm to personnel. Compliance with Personal Protective Equipment (PPE) regulations is mandatory under global safety frameworks, yet enforcing these policies in real time across vast operational floors is a major challenge.

Manual supervision suffers from critical shortcomings, including limited visual coverage, delay in incident response, subjectivity in compliance logging, and high recurring labor costs. Furthermore, high-velocity machinery and heavy equipment operate continuously, meaning even a temporary lapse in PPE compliance can lead to fatal consequences. An automated computer vision solution is required to provide zero-latency monitoring, objective spatial tracking, and instant threat escalation.

Proposed Solution

The proposed platform utilizes a state-of-the-art YOLO (You Only Look Once) architecture fine-tuned on a custom multi-class industrial safety dataset. To process high-resolution streams at scale, the deep learning model is compiled using ONNX Runtime and TensorRT optimizations, achieving low-latency edge performance on modern hardware accelerators.

The system architecture consists of four primary modules: Frame Ingestion & Preprocessing, Object Detection & Spatial Hierarchy Association, Geofencing & Intrusion Analysis, and Alert Generation & Visual Analytics Dashboard. The spatial hierarchy association module uses dynamic bounding box Intersection-over-Union (IoU) algorithms to correlate detected safety equipment directly with individual human bounding boxes, preventing false positives caused by stationary equipment resting on tables or racks.

Expected Outcomes

The deployment of this Final Year Project delivers a fully functional, production-ready computer vision solution capable of multi-camera stream analysis at 30+ frames per second per stream. The expected outcomes include:

1. High accuracy (mAP@50 > 92%) in detecting tiny and partially occluded safety items such as safety glasses, earplugs, hard hats, and vests under diverse industrial lighting conditions.
2. Reduced incident detection response time from minutes to milliseconds through real-time WebSocket alert dispatching.
3. Complete mitigation of manual safety inspection gaps by offering 24/7 continuous visual audit logs.
4. A fully customized interactive web dashboard displaying real-time compliance percentages, camera-wise violation heatmaps, and downloadable compliance reports.

Research Papers & References

1. Real-Time Personal Protective Equipment Compliance Monitoring Using Advanced YOLO Architecture
2. Deep Learning for Vision-Based PPE Detection in Construction and Industrial Environments
3. Multi-Object Tracking and Spatial Proximity Analysis for Workplace Safety Automation
4. Edge Computing Architecture for Industrial Visual Anomaly Detection