Depthai1

[tqmsh]

Migrate OAK-D pipeline to DepthAI 3.x (Pipeline 2.0)

Reference docs

• DepthAI V2→V3 Porting Guide
• V3 ObjectTracker example

Summary

• Rewrote the full on-device pipeline from depthai 2.x (legacy XLink API) to depthai 3.x (Pipeline 2.0)
• Replaced YoloSpatialDetectionNetwork + custom blob with SpatialDetectionNetwork.build() using Luxonis model zoo (yolov6-nano)
• Replaced MonoCamera/ColorCamera with Camera.build(CAM_A/B/C)
• Removed all XLinkOut nodes — use createOutputQueue() directly on node outputs
• Replaced dai.Device(pipeline) with dai.Pipeline() context manager + pipeline.start()
• Deleted scripts/download_model.py (dead code — model auto-downloads from zoo on first run)

Hardware-tested on OAK-D (MyriadX)

Setup requirements:

• USB-C ↔ USB-C cable (not USB-A → USB-C adapter — insufficient power, OAK-D draws up to 1A)
• Windows native Python (not WSL2 — usbipd-win can't reattach fast enough after OAK-D's mid-boot USB reset)
• $env:DEPTHAI_WATCHDOG=0 before running (disables MyriadX watchdog that kills device during model load)
• Always unplug/replug USB-C before running if previous session didn't exit cleanly (Ctrl+C, crash). With watchdog disabled, device can't self-reset — stale XLink state causes hangs on next run.
• If you previously used usbipd bind, run usbipd unbind --busid <id> first to restore the native Windows USB driver

Run commands:

cd C:\projects\ML-CV-Target-Tracking
$env:DEPTHAI_WATCHDOG=0; venv_win\Scripts\python main_2025.py

Proof — terminal output from hardware test (2026-04-01):

[14442C108175D2D200] [1.2] [host] [warning] Watchdog disabled! In case of unclean exit, the device needs reset or power-cycle for next run
Target ID 0: xyz=(-363mm, -350mm, 1984mm)  bbox=(130, 28, 237, 229)
Target ID 2: xyz=(322mm, -43mm, 588mm)  bbox=(417, 44, 511, 281)
Target ID 0: xyz=(-388mm, -197mm, 1928mm)  bbox=(130, 35, 227, 256)
Target ID 0: xyz=(-370mm, -200mm, 1936mm)  bbox=(130, 34, 229, 265)
Target ID 0: xyz=(-359mm, -321mm, 1948mm)  bbox=(131, 30, 234, 237)
Target ID 2: xyz=(315mm, -56mm, 574mm)  bbox=(418, 28, 512, 279)
Target ID 0: xyz=(-358mm, -337mm, 1968mm)  bbox=(131, 29, 237, 220)
Target ID 2: xyz=(313mm, -39mm, 573mm)  bbox=(417, 46, 511, 280)

• Target ID 0: person ~2m away (z≈1968mm)
• Target ID 2: person ~0.6m away (z≈573mm)
• cv2 window shows live RGB feed with green bounding boxes + distance labels

Key debugging findings

Problem	Root cause	Fix
WSL2 X_LINK_DEVICE_NOT_FOUND	usbipd reattach latency (8s) > depthai boot timeout after OAK-D USB reset	Run on Windows natively
Device hangs after watchdog warning	Stale XLink state from previous unclean exit (XLINK_WRITE_RESP 1)	Unplug USB-C, wait 10s, replug
No LED, device unresponsive	usbipd stub driver still installed (STATE=Shared)	usbipd unbind --busid <id>
YoloSpatialDetectionNetwork AttributeError	Removed in depthai 3.x	Use SpatialDetectionNetwork.build() with model zoo
depthai 2.x full pipeline always hangs	Blob upload + neural engine init causes device crash/stale state cycle	Migrated to 3.x — model zoo handles blob management

Files changed

• main_2025.py — 3.x pipeline API (dai.Pipeline(), createOutputQueue(), pipeline.isRunning())
• modules/target_tracking/stereo_node.py — Camera.build(CAM_B/C) + requestOutput((640,400))
• modules/target_tracking/spatial_detection_node.py — SpatialDetectionNetwork.build(cam, stereo, "yolov6-nano")
• modules/target_tracking/object_tracker_node.py — type hint update
• config.yaml — model_name: "yolov6-nano" replaces model_path
• requirements-pytorch.txt — depthai>=3.5.0
• .gitignore — added venv_win/, models/, .cache/, *.txt
• Deleted scripts/download_model.py (dead code)