Edge Vision Gateway

A privacy-first edge AI system that processes video feeds locally, extracts human pose data, and makes sure zero raw footage ever leaves the edge device. Designed for GDPR-compliant deployments in sensitive environments.

Privacy-Preserving Design

The core idea is simple: instead of sending raw video anywhere, the pipeline:

1. Detects humans using YOLOv11-Pose
2. Extracts 17-keypoint pose skeletons
3. Anonymizes by replacing raw pixels with skeleton overlays
4. Transmits only pose metadata + anonymized frames

Raw video never leaves the edge. Only derived, non-identifiable data gets sent out.

Performance Optimization

• OpenVINO runtime for Intel hardware acceleration
• INT8 Quantization giving a 4.2x inference speedup with minimal accuracy loss
• ONNX export from PyTorch for cross-platform deployment
• Batch processing with dynamic batch sizing based on available compute

Chain of Custody

Every processed frame includes:

• SHA-256 hash of the original frame
• Timestamp with millisecond precision
• Processing metadata (model version, confidence thresholds)
• Immutable audit log for forensic verification

Architecture

Camera Feed -> Frame Buffer -> YOLOv11-Pose -> Pose Extraction
                                                    |
                                         Skeleton Rendering
                                                    |
                                    Flask API -> Anonymized Output

Technologies

• YOLOv11-Pose: Human pose estimation
• OpenVINO: Intel edge inference runtime
• ONNX: Model interchange format
• Flask: Lightweight API server
• Python: Core pipeline
• Docker: Edge deployment container

Links

View on GitHub