ipro7_uvc_cam¶

USB 網路攝影機：OV5640、ISP、硬體 MJPEG、USB Video Class；另一個 Reef 設定在 NPU 上跑剪刀石頭布並畫進影像。教學見 USB 網路攝影機與剪刀石頭布

在 Reef 上執行¶

SW7 必須在 CAM。

README¶

A USB Video Class camera: DVP sensor → ISP → hardware MJPEG → isochronous endpoint at USB Full Speed. Hosts open it with their built-in UVC driver; nothing to install (verified on macOS). The NPU works on a second ISP channel at the same time, and what it finds is drawn into the stream the host receives.

Build¶

make -C apps/multimedia/ipro7_uvc_cam                                  # EVB    -> build/build_out/ipro7_uvc_cam_IPRO7.bin
make -C apps/multimedia/ipro7_uvc_cam DEFCONFIG=defconfig_reef_ov5640  # Reef   -> build_reef_ov5640/build_out/
make -C apps/multimedia/ipro7_uvc_cam DEFCONFIG=defconfig_reef_rps     # Reef   -> build_reef_rps/build_out/

Flash (IPRO7AI Reef EVK)¶

Board setup (jumpers, power, console) is in the IPRO7AI Reef user guide; in short, power is USB-C only. Flash with IntelProCube's command-line tool from the IntelProCube directory, using its standard 2 MB table:

ipro_iot_tool_macos --chipname ipro7 --chip-variant IPRO7AI --port /dev/cu.usbserial-XXXX \
    --baudrate 921600 --pt chips/ipro7/partition/partition_cfg_2M.toml \
    --firmware <SDK>/apps/multimedia/ipro7_uvc_cam/build_reef_ov5640/build_out/ipro7_uvc_cam_IPRO7.bin

The tool exits 0 even when flashing fails; look for [All Success]. Console: 115200 8N1 on the U6 SH1.0 wafer (the EVB build uses 921600). Download mode by hand: hold SW1 (BOOT), tap SW2 (RESET).

Image Teach¶

Image Teach (a classifier trained in IPRO AI Studio or the browser, deployed over USB) is the image mode of the Edge AI firmware apps/ai/ipro_reef_teach; this app has no Teach build. Its camera pipeline (uvc.c, uvc_isp.c, uvc_draw.c, uvc_iq_sc101iot.c) is built from apps/ai/ipro_reef_teach/image/, the same files for both.

Rock-paper-scissors build (defconfig_reef_rps)¶

The classifier, game and model of apps/ai/ipro7ai_evk_rps (full IPRO SDK) on the Reef, drawn into the stream instead of on an LCD: a play box, the live label, a 3-2-1 countdown, both moves and the score. The model is linked into the firmware, so there is nothing else to flash. Shell: rps [reset|go|mirror|print|conf|dump].

Plain webcam build (defconfig_reef_ov5640)¶

Two frame sizes, the host picks: 1280 × 720 (about 15 fps for a busy scene, limited by Full-Speed USB) and 640 × 360 (30 fps). Sensor tuning from the shell: cam rd <reg> [n], cam wr <reg> <hexbytes>, cam grid [cols rows].

EVB build (.config)¶

IPRO7 EVB with the SC101IOT sensor, 320 × 240 preview (the detector's buffers leave too little PSRAM for more). Channel 1 feeds a YOLOv8-family detector on the NPU and its boxes are drawn on the preview. The detector is chosen at compile time in uvc_ai.c:

The model is read from the mfg partition at +0x1000, so it is flashed separately from the firmware. The person + pet model is apps/platform/ipro7_demo/demo/ai/ai_model/person_pet_detect.inm.

Files¶