FPT AI Vision Agents overview
FPT AI Vision Agents is an intelligent video analytics platform built by FPT AI – FPT Smart Cloud. It helps businesses extract value from their existing camera systems to automate monitoring, analyze behavior, and optimize operations. The platform runs AI models on FPT Cloud's GPU infrastructure to process video streams in real time or on a scheduled basis.
The platform uses a microservice architecture with multi-tenant support, scaling flexibly from a handful of cameras to hundreds across multiple locations.
Key concepts
| Concept | Meaning |
|---|---|
| Organization | The top-level management unit, representing a business or store chain. Contains multiple Locations. |
| Location | A physical site (restaurant, store, office). Linked to a VMS device code to sync cameras. |
| Camera | An IP camera device managed by the system. Belongs to a Location and can have multiple Task AI assigned. |
| Task (Task AI) | A specific analysis configuration assigned to a camera (e.g. people counting, meeting check). Has its own schedule, ROI, and data source. |
| Event | An analysis result produced by the AI Engine (face detection, meeting detection, etc.). |
| Alert | A notification triggered when the system detects something needing attention (camera disconnected, poor stream quality, etc.). |
| ROI (Region of Interest) | An area drawn on the camera frame that limits what the AI analyzes. |
| VMS | A video management system (FPT's or a third party's) that provides the camera list and stream URLs to the platform. |
Key features
- Camera alerts and monitoring: detects freeze/black screen, runs periodic camera health diagnostics, pushes real-time alerts via SSE.
- Reports and analytics: customer traffic, customer presence, and shift meeting check reports — all exportable to CSV.
- Recording and storage: scheduled recording, stored on Object Storage (H.265-encoded).
Deployed use cases
- Customer counting and tracking: face recognition counts foot traffic in/out and tracks presence duration.
- Shift meeting check: uses a VLM to classify meetings as on-time / late / not held, and stores evidence images.
- Restricted zone check: detects intrusion into restricted areas and sends real-time alerts.
System architecture
The system consists of the following components, coordinating through a message broker for asynchronous processing:
| Component | Role |
|---|---|
| Platform Backend | REST API managing organizations, locations, cameras, tasks, events, and alerts; coordinates data flow, authentication, and authorization. |
| Platform Frontend | The admin web interface (Camera AI Portal) — see Camera AI Portal overview. |
| Stream Gateway | Handles camera streams (up to 32 streams per instance), detects stream quality issues, supports horizontal scaling. |
| Recording Worker | Records from the live stream or from stored video. |
| DeepStream Engine | The GPU AI pipeline, handling multi-stream processing and generating metadata for reports and alerts. |
| Video Event Worker | Handles event detection through a pipeline: chunk video → extract frames → AI inference → rules engine. |
Platform Backend schedules recording → Recording Worker records and saves segments to storage → Platform Backend queues segments for analysis → an AI Worker (DeepStream or Video Event) processes them and returns results to Platform Backend.
Next steps
- New to Vision Agents? See Getting started with Vision Agents.
- Already have an account? See Camera AI Portal overview to get familiar with the interface.
- Need quick help? See the FAQs.