Adds a community-supported hardware detector for the Qualcomm Hexagon NPU
on QCS6490 SoCs (e.g. Radxa Dragon Q6A) via QAIRT 2.37.1 / qai_appbuilder.
Mirrors the existing community-board pattern (Rockchip / Synaptics):
- frigate/detectors/plugins/qnn.py: detector plugin using yolo-generic model
type, lazy SDK import, runs pre-compiled QNN context binaries from
Qualcomm AI Hub.
- docker/qcs6490/: Dockerfile (two-stage; rebuilds qai_appbuilder wheel
inside Frigate's image to match libstdc++ ABI), bake target (qcs6490.hcl),
make targets (qcs6490.mk), and a host-side user_installation.sh that
installs fastrpc, the QCS6490 firmware (cDSP image + skel libs), and
configures cdsprpcd.
- .github/workflows/ci.yml: qcs6490_build job mirroring synaptics_build.
- CODEOWNERS: /docker/qcs6490/ + qnn.py.
- docs: new "Qualcomm Hexagon NPU" sections under Community Supported
Detectors in object_detectors.md, plus matching entries in
installation.md and hardware.md.
Performance on Radxa Dragon Q6A (Hexagon v68, ~12 TOPS), YOLOv8n 640x640:
~10ms per inference under light load, ~24ms with 5 RTSP cameras live.
Closes#18602.
