AWS Architecture Blog·April 1, 2026Designing a Scalable Computer Vision System for Workplace Safety Monitoring on AWS
This article outlines a serverless, event-driven architecture for a computer vision and generative AI-based workplace safety monitoring system. It details how to scale the solution to hundreds of sites and thousands of cameras, focusing on image collection, anonymization, training pipelines, and inference. Key system design considerations include data privacy, security, and continuous model improvement through feedback loops.
Read original on AWS Architecture BlogThe article presents a robust, scalable architecture for automated workplace safety monitoring using computer vision and generative AI. This system addresses the limitations of manual safety audits by providing continuous, real-time oversight of PPE compliance and zone-based hazard detection across numerous facilities. It emphasizes a serverless, event-driven approach designed for efficiency and high scalability.
Core Architectural Principles
- Serverless and Event-Driven: The architecture leverages AWS Lambda, Amazon S3, AWS Step Functions, and Amazon EventBridge to build a highly scalable and cost-effective system. This allows for processing massive volumes of image data from thousands of cameras efficiently.
- Distributed Across AWS Accounts: To ensure proper security, operational segregation, and data isolation, the solution is distributed across multiple AWS accounts. This includes separate accounts for training pipelines, image collection, the end-user web application, and analytics.
- Privacy-Preserving: Human faces and identifiable features are blurred using Amazon Rekognition and custom Python code immediately after image capture, before images are replicated or used for training/inference, to protect Personally Identifiable Information (PII).
Key System Components and Data Flows
The system's workflow begins with image collection from site cameras, stored temporarily in a restricted Amazon S3 bucket for anonymization. These anonymized images are then replicated to S3 buckets across different accounts for training, inference, and the web application. This multi-stage data handling is critical for both privacy and operational efficiency.
ML Training Pipeline
- Ground Truth Generation: AWS Step Functions orchestrate Amazon SageMaker Ground Truth labeling jobs, triggered by Amazon EventBridge. This workflow integrates Zone User feedback and saved ML model predictions to prioritize data labeling for underperforming classes and cameras.
- Data Processing and Storage: AWS Lambda transforms completed labeling jobs into a suitable format, storing metadata in Amazon DynamoDB and annotations in an S3 bucket.
- Model Building and Promotion: Amazon SageMaker AI Pipelines execute model training workflows. Approved models trigger an EventBridge event, which a Lambda uses to update the SageMaker AI endpoint via a CI/CD pipeline, decoupling model science from application deployments.
Inference Pipeline
Each safety use case has its own inference pipeline. Once an anonymized image lands in its dedicated S3 bucket, it triggers an Amazon SNS notification, initiating the hazard detection process. This continuous monitoring acts as a digital safety supervisor, distinguishing normal workflows from potential hazards.
Scalability and Maintainability
The design leverages serverless services for automatic scaling and manages model updates through a well-defined CI/CD process. Decoupling the model training and deployment cycles ensures that data scientists can iterate on models without directly impacting the application code, and engineers can manage infrastructure changes independently.