Real-time on-site inference
Sub-second detection and alarm generation on live streams — no cloud round-trip, no dependency on internet uptime.
Turn passive surveillance into active operational intelligence. VisionX analyses live video locally on purpose-built NPU hardware, turning detections into immediate, structured events.

System overview
Instead of streaming everything to a cloud GPU cluster, inference runs directly on dedicated AI silicon inside the camera, the recorder or the on-site appliance. Operating cost stays flat as channels grow.
Sub-second detection and alarm generation on live streams — no cloud round-trip, no dependency on internet uptime.
Frames are analysed where captured. Only metadata and selected evidence clips are shared — simplifying compliance.
Purpose-built NPUs replace power-hungry GPGPUs — cutting energy, cooling and rack cost across the lifecycle.
Train, quantize and deploy custom detectors. Standard ONNX converts to optimized .dxnn or .sima NPU binaries.
Pipeline
IP, thermal and PTZ cameras feed H.264/H.265 streams into the VisionX layer.
DEEPX and SiMa.ai NPUs run detection, classification and tracking on every frame.
Zones, schedules and thresholds turn detections into structured events.
Alarms to VMS, mobile, e-mail and relays; metadata indexed for forensic search.
Analytics library
Every function runs on-device in real time, combines with zones and schedules, and outputs structured events to the VMS, dashboards and third-party systems.
Virtual line crossing with direction control — gates, fence lines, restricted corridors.
Two lines crossed in defined order and time window — near-zero nuisance alerts.
Polygonal zone monitored for entry, exit and stay events with dwell threshold.
People in restricted zones outside authorised schedules — instant alert to operator, mobile and VMS.
Classifies vehicle types — car, bus, truck, motorcycle, rickshaw — enabling type-specific alerts.
Reads and matches vehicle plates (Bengali + English) for access control, enforcement and watchlists.
Newly placed objects escalated if they remain beyond a set duration.
Alarm when a monitored asset disappears from the scene beyond configured time.
Movement faster than a configured speed inside a zone — flight or reckless driving.
Deploy anywhere
Three deployment paths — the same software layer, the same analytics library, the same integration APIs. Mix all three across one unified console.
Comparison
| VisionX on edge NPU | Cloud AI service | On-prem GPGPU server | |
|---|---|---|---|
| Alert latency | < 1 second, on-site | Round-trip dependent | Sub-second, high power draw |
| Power per engine | 1–10 W | Off-site (billed) | 40 W+ per accelerator |
| Video leaving site | None — metadata only | Full streams uploaded | None |
| Internet dependency | None | Continuous uplink required | None |
| Cost as channels grow | Flat — capex per card | Per-stream recurring fees | Rack, cooling and energy scale |
| Data sovereignty | Full | Provider jurisdiction | Full |
Applications
PPE and helmet compliance, restricted-zone intrusion, fall detection on production floors, unattended machinery posts, forklift/pedestrian conflict zones.
Perimeter protection after hours, loading-bay dwell time, abandoned pallets, vehicle parking enforcement and object-missing detection.
Footfall counting, queue and dwell heatmaps, loitering at high-value fixtures, loss-prevention tripwires, after-hours intrusion.
Fence-line double tripwire, unauthorized-person detection, camera tamper alerts, unattended object detection, thermal monitoring.
Deployment enquiry
Book a live demo, request a dev kit or size a deployment with our engineering team.