Logo Event-Driven Architecture & Industrial Edge EN ES

NeuroFace CV — Computer Vision Pipeline

Overview

The NeuroFace Computer Vision pipeline extends the demo’s event-driven architecture, integrating face recognition with event streaming via Kafka and email notifications.

Components

Component Function Namespace

NeuroFace Frontend

Angular app for face registration and live detection via webcam (WebRTC)

neuroface

NeuroFace Backend

FastAPI + OpenCV API with LBPH recognizer — trains labels and exposes /api/labels

neuroface

OpenVINO Model Server (OVMS)

Intel inference server serving face-detection-retail-0005 (FP16) via REST/gRPC, deployed as KServe InferenceService

neuroface

Model PVC

1 Gi PersistentVolumeClaim storing Intel Model Zoo .xml/.bin files

neuroface

Jupyter Workbench

OpenShift AI notebook for model experimentation

neuroface

CV Processor

Python microservice (UBI9) — polls OVMS /v1/config and Backend /api/labels every 30 s

kafka-cdc

Kafka Topic

cv.face.detections on Strimzi cdc-cluster (SASL_SSL + SCRAM-SHA-512)

kafka-cdc

Kafka Consumer

Same Python process — consumes events and routes by event_type

kafka-cdc

Mailpit

SMTP + HTTP API — receives HTML emails formatted by event type

openshift-lightspeed

Event Flow

The CV Processor runs a continuous polling loop:

  1. Register face — The user registers a face in the NeuroFace UI (webcam → Frontend → Backend LBPH)

  2. Poll OVMS — GET /v1/config every 30 seconds to check model status

  3. Poll Backend — GET /api/labels every 30 seconds to detect newly registered faces

  4. Produce to Kafka — Publishes JSON events to topic cv.face.detections (SASL_SSL authenticated)

  5. Consume from Kafka — The same process reads events and routes by event_type

  6. Send email — Formats HTML and sends POST /api/v1/send to Mailpit

Event Types

Type Description Recipient

person_registered

A new label is detected in the /api/labels response

Security Admin — admin@neuralbank.io

ovms_model_status

OVMS /v1/config returns the model configuration

ML Ops — mlops@neuralbank.io

Kafka Event Schema

{
  "event_type": "person_registered",
  "timestamp": "2025-06-15T14:32:00Z",
  "source": "neuroface-lbph",
  "person_name": "Max",
  "face_count": 1,
  "detection_method": "lbph",
  "camera_id": "label-registry"
}
{
  "event_type": "ovms_model_status",
  "timestamp": "2025-06-15T14:32:30Z",
  "source": "openvino-model-server",
  "model_name": "face-detection-retail-0005",
  "model_server": "ovms",
  "ovms_response": { "face-detection-retail-0005": { "model_version_status": [{"state": "AVAILABLE"}] } },
  "camera_id": "ovms-monitor"
}

PPE Safety Detection (YOLO + OpenShift AI + Kafka)

The PPE Safety Detection module extends the NeuroFace platform with real-time Personal Protective Equipment monitoring. Webcam frames are sent to a YOLOv8n serving endpoint for object detection, and a Granite 3.1 LLM generates natural-language safety analysis when violations are detected. Events are published to Kafka for compliance dashboards.

PPE Architecture

Component Role Namespace

YOLO PPE Serving

Python HTTP server running Ultralytics YOLOv8n — receives JPEG, returns bounding boxes and class labels

neuroface

Granite 3.1 2B LLM

KServe InferenceService — generates safety analysis on PPE violations

neuroface

Kafka Topic

cv.ppe.detections on Strimzi cdc-cluster with SASL_SSL + SCRAM-SHA-512

kafka-cdc

Kafka DLQ

dlq.ppe-errors — dead letter queue for failed processing

kafka-cdc

Jupyter Workbench

ppe-yolo-detection.ipynb — consumes Kafka events, plots compliance charts

neuroface

PPE Data Flow

Each webcam frame follows this path:

  1. Capture — Angular frontend captures a frame via WebRTC

  2. YOLO Inference — Backend forwards to YOLO serving (/v1/predict), gets bounding boxes with class names and confidence scores

  3. Compliance Check — Backend compares detections against expected PPE classes (hardhat, safety-vest, goggles) and assigns status: compliant, violation, or no_persons

  4. LLM Analysis (on violation) — Granite 3.1 generates a safety analysis explaining risks

  5. Kafka Publish — Full detection event published to cv.ppe.detections

  6. Notebook Consumption — Jupyter renders compliance dashboards with matplotlib

PPE Kafka Event

{
  "event_type": "ppe_compliance_check",
  "source": "neuroface-ppe-backend",
  "timestamp": "2026-04-19T18:30:00Z",
  "ppe_status": "violation",
  "person_count": 1,
  "expected_ppe": ["hardhat", "safety-vest", "goggles"],
  "present_ppe": ["safety-vest"],
  "missing_ppe": ["hardhat", "goggles"],
  "detections": [
    {"class_name": "person", "confidence": 0.92, "bbox": {"x1": 120, "y1": 50, "x2": 380, "y2": 470}},
    {"class_name": "safety-vest", "confidence": 0.87, "bbox": {"x1": 145, "y1": 180, "x2": 355, "y2": 390}}
  ],
  "llm_analysis": "PPE violation: 1 operator missing hardhat and goggles. Risk: head injury and eye exposure. Immediate corrective action required."
}

Security

  • Kafka: SASL_SSL with SCRAM-SHA-512 (credentials cdc-user from Kubernetes Secret)

  • Kafka Console OIDC: Authenticated via Keycloak (RHBK) — roles: platformadmin → admin, userN → viewer

  • OVMS / YOLO: Cluster-internal access only (no external Route)

  • LLM endpoint: KServe with ServiceAccount token injection

  • Inter-namespace traffic: All Kafka connections use TLS

Integration with OpenShift AI

The NeuroFace model is served through OpenShift AI using KServe:

Resource Detail

ServingRuntime

ovms-face-detection — custom runtime based on the OVMS image

InferenceService

face-detection-retail-0005 — RawDeployment mode, storage from PVC

Jupyter Workbench

neuroface-workbench — for model experimentation and debugging

GitOps and Argo CD

All CV pipeline resources are deployed via GitOps:

  • field-content-neuroface-cv: Argo CD Application for OVMS, InferenceService, CV Processor, Notebook

  • field-content-mailpit: Argo CD Application for Mailpit (SMTP sink)

  • field-content-kafka-console: Console with OIDC enabled

The full definition is in examples/helm/components/neuroface-cv/templates/all.yaml.

Verification

OVMS Model Status

curl -s http://face-detection-retail-0005-predictor.neuroface.svc.cluster.local:8080/v1/config

Registered Labels

curl -s http://neuroface-backend.neuroface.svc.cluster.local:8000/api/labels

Events in Kafka

Go to Kafka Console → Topics → cv.face.detections → Messages to view events in real time.

Notification Emails

Open Mailpit (https://n8n-mailpit-openshift-lightspeed.) to see emails sent by the consumer.

Access

Service URL

NeuroFace App

https://neuroface-neuroface.

OpenShift AI

https://data-science-gateway./projects/neuroface

Kafka Console

https://kafka-console-kafka-cdc.

Mailpit

https://n8n-mailpit-openshift-lightspeed.

Screenshots

NeuroFace — Face Registration UI
NeuroFace — Detection Pipeline
NeuroFace — Kafka Events
NeuroFace — Mailpit Notifications