Skip to main content

RZ/V2N DRP-AI3 YOLOv3 Object Detection

View source on GitHub

This guide walks through building and running the YOLOv3 object detection reference on the SolidRun RZ/V2N HummingBoard. The app pulls frames from a looping video file (default) or the IMX678 MIPI CSI-2 camera, runs YOLOv3 (Darknet/COCO) inference on the RZ/V2N's on-chip DRP-AI3 accelerator using Renesas's pre-compiled TVM bundle from the RZ/V2N AI SDK v6.30, and renders the annotated feed full-screen on Wayland/Weston.

Prerequisites​

  • macOS 10.12+ or Linux (Ubuntu 22.04+, Fedora 39+)
  • Docker Desktop installed and running
  • The latest Avocado CLI
  • SolidRun RZ/V2N HummingBoard SoM
  • HDMI display connected to the carrier
  • microSD card or eMMC for boot media
  • curl on the build host (used by fetch-model.sh and fetch-video.sh)

Initialize​

avocado init --reference rzv2n-drpai-yolo rzv2n-drpai-yolo
cd rzv2n-drpai-yolo

Fetch the DRP-AI3 model bundle​

./fetch-model.sh

Pulls Renesas's pre-compiled YOLOv3 bundle from rzv_ai_sdk v6.00 — the small graph metadata + DRP-AI/AI-MAC engine descriptors from the repo, and the multi-MB compiled deploy.so from the release asset. Lands at app/overlay/usr/lib/rzv2n-drpai-yolo/model/yolov3/. .gitignored.

Fetch the sample video​

./fetch-video.sh

Default: a 15s / 8 MB / 1080p H.264 clip of a busy NYC sidewalk — pedestrians, UPS truck, yellow taxi. Multiple YOLO targets per frame, the kind of scene a battery-powered outdoor camera (Ring / Nest / Arlo) would capture. Pexels free-license.

To use a different clip:

VIDEO_URL=https://your.cdn/clip.mp4 ./fetch-video.sh

The script's comments list a couple of curated alternatives (front-door delivery, suburban approach). The fetched file lands at app/overlay/usr/lib/rzv2n-drpai-yolo/sample.mp4 — .gitignored.

Install​

avocado install -f

Pulls the SDK container image and resolves runtime dependencies — lib-tvm (DRP-AI3 inference runtime), drpai (UAPI header for /dev/drpai0), the Renesas memory manager userspace libraries, GStreamer with libav/qtdemux/h264parse for video decode + waylandsink, OpenCV, and Weston.

Build​

avocado build

app-compile.sh clones rzv_drp-ai_tvm at v2.5.1, runs setup/make_drp_env.sh with PRODUCT=V2N to overlay Renesas's DRP-AI-patched TVM headers (kDLDrpAi), then cmake cross-compiles the YOLOv3 inference binary linking against the target's libtvm_runtime.so.

Deploy​

avocado provision -r dev --profile sd --env AVOCADO_SD_DEVICE=/dev/sdX

(Replace /dev/sdX with your SD card device. Use --profile emmc for eMMC.)

Verify​

Log in as root with an empty password. The app service starts automatically after Weston:

systemctl status rzv2n-drpai-yolo
journalctl -u rzv2n-drpai-yolo -f

Expected output:

rzv2n-drpai-yolo starting
  source: video file /usr/lib/rzv2n-drpai-yolo/sample.mp4
  model: /usr/lib/rzv2n-drpai-yolo/model/yolov3
  drp_start_addr: 0xd0000000
  model loaded — outputs=3
  pipelines running
  frames=150 inference_avg=78.4ms detections_last=2

The annotated video plays full-screen on the connected HDMI display, looping when it reaches end-of-file.

Customize​

Use a different video​

Edit Environment=VIDEO_PATH=... in app/overlay/usr/lib/systemd/system/rzv2n-drpai-yolo.service, or drop a new file at /usr/lib/rzv2n-drpai-yolo/sample.mp4. Any container/codec GStreamer's decodebin can handle works (MP4, MKV, AVI; H.264, H.265, VP9, etc.). Re-run avocado build && avocado provision after changing.

Switch to the IMX678 camera (currently broken)​

The camera path is wired but disabled by default because the SolidRun rzg2l-cru driver misreports its V4L2 format on RZ/V2N at 4K — frames decode as colored vertical stripes. To experiment anyway:

  1. Empty VIDEO_PATH in the service (Environment=VIDEO_PATH=).
  2. Enable the camera-init unit: systemctl enable --now rzv2n-drpai-yolo-camera.service.
  3. The app falls into the v4l2src branch, capturing from /dev/video0.

A proper fix would either patch the kernel to truncate 12→8 bit correctly, or change main.cpp to read the buffer as 1920×2160 16-bit-BE Bayer and demosaic via cv::cvtColor(..., COLOR_BayerBG2BGR).

Adjust detection thresholds​

Environment=CONFIDENCE_THRESHOLD=0.4
Environment=NMS_THRESHOLD=0.5

Override the DRP-AI reserved-memory base​

If the drp_reserved carveout in your DTSI sits elsewhere:

Environment=DRP_START_ADDR=0x90000000

Rebuild after changes​

avocado build
avocado provision -r dev --profile sd --env AVOCADO_SD_DEVICE=/dev/sdX