Multi-Sensor Perimeter Perception for a Critical Site

The operational problem

Perimeter security at high-consequence sites usually fails in the perception layer rather than the model layer. Camera-only pipelines lose targets in fog, glare, and darkness; unsynchronized sensor rigs produce tracks that cannot be reconciled; and any pipeline that depends on a cloud endpoint is disqualified outright at sites with no permitted egress.

What Enfuse built

• Time-synchronized capture across GMSL2 cameras, spinning LiDAR, and thermal imagers
• Intrinsic and extrinsic calibration workflow with a repeatable field re-calibration procedure
• Edge detection and classification on Jetson with TensorRT-optimized models
• Late-fusion tracker producing a single object track per physical entity across modalities
• Spatial grounding of tracks into site coordinates against a surveyed 3D model
• Signed offline model and configuration bundles for disconnected updates

Architecture

• See — synchronized multi-modal capture at the fence line
• Understand — per-sensor detection and classification on the edge device
• Locate — late fusion and tracking, then projection into site coordinates
• Predict — trajectory and dwell-time analysis for approach and loiter behavior
• Act — alerting into the existing operations console and access-control workflow

Hardware and software

• NVIDIA Jetson AGX Orin / Orin NX (edge inference)
• NVIDIA B200 or H200 class server (training, re-identification, digital twin)
• ROS 2 / Isaac ROS, DeepStream, TensorRT
• Ouster or Hesai LiDAR, GMSL2 camera arrays, thermal imagers

Deployment environment

On-premises, edge-deployed, designed to run fully air-gapped.

Results

• Sub-50ms edge inference budget per sensor node
• Continuous operation with no outbound network dependency
• Single fused track per entity instead of per-camera detections

Reference architecture. Figures are engineering design targets for this pattern, not measured results from a named client deployment.