Transforming Industrial Safety Monitoring

Detecting industrial safety risks in seconds instead of minutes is critical for keeping workers safe on an active plant floor. To address this challenge, Tata Elxsi built IRIS, a real-time industrial safety platform on AWS. As a global provider of design and technology services, Tata Elxsi combines engineering depth with artificial intelligence and computer vision to help enterprises modernize safety-critical physical operations. The IRIS platform was specifically designed for organizations that operate extensive camera infrastructure but struggle to turn passive video feeds into real-time, actionable intelligence.

Manufacturing plants, warehouses, logistics hubs, and chemical facilities frequently operate hundreds or thousands of cameras covering production lines, hazardous zones, and loading areas. However, closed-circuit television typically functions as a recording system rather than a prevention system, meaning incidents usually surface only after they occur. Furthermore, human control-room operators face natural monitoring limits, often taking between 15 and 45 minutes to detect unsafe conditions.

Edge Architecture and Video Filtering

The IRIS platform operates as a serverless, event-driven pipeline that observes activity at the edge before sending relevant metadata to the cloud. The workflow begins with a dedicated edge-compute tier deployed using AWS IoT Greengrass on industrial-grade, GPU-equipped edge servers connected to the facility camera network over RTSP or ONVIF protocols.

At the edge, IRIS extracts video frames at a configurable rate of two to five frames per second and applies motion-based filtering alongside a lightweight first-pass model. Frames that pass these filters are uploaded to Amazon Simple Storage Service (Amazon S3). This decoupling of the image path from the metadata path keeps streaming events lightweight, while filtering at the edge successfully reduces the volume of frames sent to the cloud by approximately 70 to 80 percent.

Real-Time Event Streaming and Processing

Following edge processing, safety-relevant metadata and events are streamed directly into Amazon Kinesis Data Streams, which functions as the real-time event backbone of the platform. The stream carries frame metadata, motion events, edge detection candidates, camera telemetry, and contextual safety information without transmitting heavy raw video files.

The stream operates in an on-demand capacity mode, automatically scaling throughput during shift changes or multi-incident bursts. In production deployments by Tata Elxsi, sustained throughput reaches between 2,000 and 5,000 events per second, with burst capacity handling up to 15,000 events per second and measured ingestion latency remaining under 200 milliseconds.

Computer Vision Intelligence

Event streams are consumed by custom computer vision models deployed on Amazon SageMaker AI, serving as the core intelligence engine for the platform. Separate real-time endpoints are provisioned for each model family so they can scale independently according to operational demands.

The deployed models cover multiple safety domains, including personal protective equipment compliance using fine-tuned YOLOv8 object detection, restricted-zone monitoring with spatial geofencing, worker safety analytics using SlowFast temporal action recognition, and vehicle proximity tracking for industrial machinery.

Sources

• AWS Machine Learning BlogHow Tata Elxsi detects industrial safety risks in seconds on AWS