AI-Powered Safety Compliance Monitoring System

1. Executive Summary

HelmNet is a computer vision solution developed to automate safety compliance monitoring in high-risk environments such as construction sites and manufacturing plants. By leveraging deep learning object detection architectures, the system identifies individuals and verifies the usage of safety helmets in real-time. This project bridges the gap between safety protocols and operational reality, reducing liability risks and enhancing worker safety through automated surveillance.

2. Business Context & Problem Statement

The Challenge

Workplace safety is a paramount concern in industries like construction, mining, and manufacturing. Despite strict regulations regarding Personal Protective Equipment (PPE), specifically safety helmets, non-compliance remains a leading cause of preventable head injuries and fatalities.

• Manual Monitoring Limitations: Traditional safety checks rely on human supervisors, which is labor-intensive, prone to error, and impossible to scale across large, complex sites.
• Operational Risk: Accidents resulting from PPE non-compliance lead to severe legal consequences, financial losses, and operational downtime.

The Objective

The goal was to develop a robust AI model capable of:

1. Detecting persons within a video feed or image.
2. Classifying whether they are wearing a safety helmet or are unprotected (head).
3. Providing a scalable, automated tool to flag safety violations instantly.

3. Technical Architecture & Methodology

Data Strategy

The project utilized a labeled dataset comprising 5,000 images, annotated with bounding boxes for three specific classes: Helmet, Head (no helmet), and Person.

• Data Augmentation: To ensure the model was robust against varying lighting conditions, angles, and occlusions common in construction sites, the albumentations library was employed. Transformations included horizontal flips, brightness/contrast adjustments, and slight rotations.
• Data Processing: A custom SafetyHelmetDataset class was implemented to handle image loading, target transformation, and integration with the PyTorch data loader.

Model Architecture

The solution was built using Transfer Learning to maximize performance with limited training data.

• Framework: PyTorch and Torchvision.
• Algorithm: Faster R-CNN (Region-based Convolutional Neural Network), a state-of-the-art architecture for object detection known for its high accuracy.
• Backbone: MobileNetV3-Large with FPN (Feature Pyramid Network). This backbone was selected specifically to balance accuracy with inference speed, making the model suitable for deployment on edge devices or mobile platforms often used in field operations.
• Anchor Generator: Custom anchor sizes and aspect ratios were configured to effectively capture objects of various scales (e.g., small helmets in the distance vs. close-ups).

Training Pipeline

• Optimizer: Stochastic Gradient Descent (SGD) with a learning rate of 0.005, momentum of 0.9, and weight decay of 0.0005 to prevent overfitting.
• Scheduler: A StepLR scheduler was implemented to decay the learning rate every 3 epochs, allowing the model to fine-tune weights effectively as training progressed.
• Environment: The model was trained using GPU acceleration (CUDA) to expedite the computational load of the convolutional layers.

4. Key Results and Evaluation

The model was evaluated using standard object detection metrics, specifically Mean Average Precision (mAP), using the COCO evaluator tool.

Performance Highlights

• Detection Capability: The model demonstrated the ability to distinguish between a "Helmet" (safe) and a "Head" (unsafe) with high confidence scores.
• Visual Validation: Inference tests on unseen data showed accurate bounding boxes drawn around workers, correctly labeling those compliant with safety gear versus those violating protocols.
• Loss Convergence: Training logs indicated a steady decrease in classifier and box regression loss, confirming the model effectively learned the spatial and feature representations of the safety gear.

5. Future Scope & Recommendations

To transition this prototype into a production-grade enterprise solution, the following enhancements are recommended:

• Real-Time Video Integration: Adapt the inference pipeline to process RTSP video streams from CCTV cameras.
• Edge Deployment: Convert the model to ONNX or TensorRT formats to run efficiently on edge devices (e.g., NVIDIA Jetson) directly at construction sites.
• Alert System: Integrate an alert API to notify site managers via SMS or dashboard notifications immediately when a "Head" (no helmet) class is detected in a restricted zone.