A compact vision module that runs detection and tracking entirely on device.
A palm-sized vision device for environments where video cannot leave the room. Detection, tracking and event classification all run locally on an NPU, and only structured events reach the cloud.
The product needed continuous multi-stream inference under 4 W, with no video egress for privacy reasons, and a bill of materials that survived a consumer price point.
We paired a quantized detection backbone with a lightweight tracker so the heavy model runs at a reduced cadence while tracking holds identity between inferences. The result is a stable 30 fps pipeline on hardware that could not brute-force it.
A quad-core SoC with an integrated 6 TOPS NPU, MIPI CSI-2 camera input, and a 6-layer board with a controlled-impedance camera interface.
An INT8 detection model deployed to the NPU, with a Kalman-based tracker and a small temporal classifier for event logic.
An embedded Linux image built with Yocto, a GStreamer-based capture pipeline, and a web console for configuration and live event review.
Hardware
AI
Embedded
Cloud