The conventional approach to AI-enabled thermal imaging requires three separate elements: an imaging module, a cable to a compute board, and an AI accelerator chip (GPU, NPU, or FPGA). Each element adds weight, power consumption, connector reliability risk, and integration complexity. The NEXUS LV0619B takes a different approach.

What “On-Device AI” Actually Means

The NEXUS LV0619B integrates a dedicated neural processing unit (NPU) directly on the same PCB as the 640×512 uncooled LWIR detector. The imaging pipeline — raw detector output → radiometric correction → NUC → AI inference — runs entirely within the module. No external processor is required for detection.

What comes out of the module:

  • Corrected 640×512 thermal video stream (MIPI or CML interface)
  • Detection results: bounding boxes, class labels, confidence scores — over UART/RS422

The host system receives both streams simultaneously. It can display or record the video, and act on detection events without performing any AI computation locally.

AI computer vision target detection on thermal imaging display
On-device AI inference eliminates the external compute board — detection results are delivered alongside video from a single compact module

Detection Capabilities

The NEXUS LV0619B ships with pre-trained models for common target categories in security and defense applications:

  • Person detection: pedestrian and prone human detection, optimized for thermal contrast in varied background clutter
  • Vehicle detection: car, truck, motorcycle — distinguishing vehicle heat signatures from background
  • Multi-class simultaneous: detect people and vehicles in the same scene concurrently

Detection performance (typical, on-module inference):

Target Class Detection Range (typical) Frame Rate Latency
Standing person 400–700 m 25 fps < 40 ms
Vehicle 600–1200 m 25 fps < 40 ms
Multi-class Multiple targets 25 fps < 45 ms

Range figures depend on lens focal length and target thermal contrast. The above uses a 25 mm f/1.0 lens under moderate thermal conditions.

Why On-Device Matters for Deployed Systems

Bandwidth reduction

A raw 640×512 thermal video stream at 30 fps requires approximately 40–80 Mbps of bandwidth depending on encoding. Detection results — bounding boxes with coordinates and classifications — require less than 10 Kbps. For systems with limited data links (long-range wireless networks, satellite, military tactical radios), this is the difference between feasibility and infeasibility.

Latency

When inference runs on the same board as the sensor, detection-to-output latency is under 40 ms — comparable to a single video frame. Offloading to a remote server adds round-trip network latency that can reach hundreds of milliseconds, degrading track quality and alert response time.

Reliability

Fewer boards mean fewer connectors, fewer power rails, and fewer software dependencies. A single-board detection system eliminates an entire failure mode category compared to a multi-board design.

Deployability in constrained platforms

For UAV gimbals under 200 g, man-portable devices, and small UGV systems, the weight and volume of an external AI compute board (typically 100–300 g with heat spreader and power supply) can be prohibitive. The NEXUS LV0619B runs detection inference at the module’s 2.5 W total power draw — no additional board, no additional power budget.

Custom Model Support

IRmodules provides the model deployment framework for customers who need to run custom-trained detection models on the NEXUS platform. Supported model formats and quantization schemes are documented in the NEXUS integration SDK. Typical custom model deployment workflow:

  1. Train model using customer dataset (PyTorch or TensorFlow)
  2. Export to ONNX format
  3. Quantize to INT8 using IRmodules’ optimization toolkit
  4. Deploy to module via update utility

This workflow enables customers to adapt the on-device detection to specific target types — maritime vessels, specific vehicle models, wildlife species — without changing the hardware.

Integration with Existing Platforms

The NEXUS LV0619B is a drop-in electrical replacement for the SPECTRA L06 (same 35×35 mm form factor, same DC 5V power, same video interface). Existing enclosures, cables, and mounting hardware are fully compatible. The addition of detection capability requires only a UART connection to the host for receiving detection results.

For new designs, the NEXUS platform represents the future of compact, deployable AI-enabled EO/IR — removing the architectural constraint that has historically forced a trade-off between AI capability and system weight.