This repository was archived by the owner on Apr 10, 2026. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathoakd.py
More file actions
1213 lines (1014 loc) · 50.1 KB
/
oakd.py
File metadata and controls
1213 lines (1014 loc) · 50.1 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
"""
ERIC — OAK-D Lite Depth Camera (Layer 1 + Layer 2)
DepthAI 3.x API — Pipeline runs itself via pipeline.start().
Layer 1 Safety (industrial-standard reactive safety):
Every depth frame is checked instantly in the reader thread:
front depth < OAKD_STOP_DIST → motors.stop() immediately
front depth < OAKD_SLOW_DIST → motors.slow() immediately
floor drop detected (high OR medium confidence) → motors.stop()
This fires in the reader thread — zero latency, independent of Cosmos.
Floor drop check runs at FLOOR_CHECK_HZ to limit CPU (numpy is expensive).
Layer 2 Person/Animal Detection (YOLOv8n on OAK-D Myriad X VPU):
Runs YOLOv8n entirely on the OAK-D's built-in Myriad X — zero Jetson RAM.
Detects: person, dog, cat, horse, sheep, cow, bird + large animals (COCO).
Every detection frame:
confidence > YOLO_MIN_CONFIDENCE
AND OAK-D spatial depth confirms distance
AND distance < YOLO_SLOW_DIST → slow motors
AND distance < YOLO_STOP_DIST → stop motors (Layer 2 motor guard)
→ fires _yolo_callback(label, distance_m, bearing_deg) for mission.py
No Cosmos involved — pure Myriad X VPU hardware detection.
Falls back gracefully if blob not found — depth pipeline still works.
Fixes applied (v3 → v4)
────────────────────────
1. ISP scale crash fix — the IMX378/214 sensor only supports 1080_P, 4_K, or
12_MP. At 1080_P it outputs native 2104×1560 (not 1920×1080). setIspScale
ratios don't land cleanly on standard sizes from this base. Fix: keep THE_1080_P
and use setVideoSize(960, 540) to explicitly set the ISP output size. This
avoids the post-proc mismatch and fits within OAK-D Lite SIPP memory budgets.
2. _avoidance_in_progress UnboundLocalError — the variable was set inside the
`else` branch of the front-obstacle check, then referenced in the floor-drop
check below (which is outside that branch). If valid_patches == 0, the floor-
drop block raised UnboundLocalError. Fixed: moved the avoidance check to the
top of _safety_check() so it is always defined before either branch uses it.
3. Reconnect watchdog race condition — _ensure_reconnect_watchdog() used a
plain boolean _reconnect_started, so if init_oakd() was called from two
threads simultaneously (e.g. user + reconnect watchdog), two watchdog threads
could start. Fixed: replaced with a threading.Lock guard.
4. _teardown() doesn't wait for reader thread — after _teardown() sets
_oakd_ok=False, the reader thread may still be mid-frame when init_oakd()
starts the new pipeline, leading to stale frame data and potential lock
contention. Fixed: _teardown() now joins the reader thread (up to 2s) before
tearing down the pipeline.
5. YOLO FPS reduced to 5 on OAK-D Lite — the Lite variant has a lower-spec
Myriad X than the full OAK-D. Running colour at 10 FPS alongside YOLO
spatial detection and stereo depth at 15 FPS overflows SIPP memory,
producing the E_OUT_OF_MEM + E_FINALISE_FAIL crash loop seen in logs.
Colour camera FPS reduced back to 5. The 3:1 ratio (15 FPS mono / 5 FPS
colour) is within the Lite's validated operating range.
6. Floor-drop void check used floor_check_interval parameter but then
re-computed 1.0/FLOOR_CHECK_HZ inline — the parameter was redundant and
confusing. Unified to use the global constant directly, parameter removed.
7. get_depth_at() patch clamp was asymmetric — upper bounds used cx+patch_px
without +1, so numpy slice [cx-10:cx+10] gives 20 pixels but
[cx-10:min(w, cx+10)] could give 19 if cx+10==w. Cosmetic but fixed for
consistency.
SLAM changes (vs oakd.py):
- Depth frames published as /oakd/depth ROS2 topic (sensor_msgs/Image)
- Nav2 costmap subscribes to this for 3D obstacle awareness
- publish_depth_for_costmap() called from _reader_loop after safety check
- Adds ~2ms overhead per frame — acceptable at 15 FPS stereo
Disconnect handling (unchanged):
- _reader_loop detects fatal XLink errors and exits cleanly.
- _oakd_ok cleared immediately so all callers see unavailable.
- Reconnect watchdog retries init_oakd() every RECONNECT_INTERVAL_S s.
- Log spam suppressed: one warning per disconnect, silence until back.
"""
import datetime
import logging
import threading
import time
import numpy as np
from pathlib import Path
log = logging.getLogger("eric.oakd")
# ── Core state ────────────────────────────────────────────────────────────────
_pipeline = None
_depth_queue = None
_depth_frame = None
_lock = threading.Lock()
_oakd_ok = False
_reader_thread = None
# ── SLAM / Nav2 depth publishing ─────────────────────────────────────────────
_depth_pub = None # ROS2 publisher for /oakd/depth (Nav2 costmap input)
_depth_pub_hz = 5.0 # publish rate — matches colour camera FPS
_last_depth_pub = 0.0 # timestamp of last depth publish
# ── Reconnect watchdog ────────────────────────────────────────────────────────
_reconnect_thread = None
_reconnect_lock = threading.Lock() # FIX 3: replaces bare boolean flag
_reconnect_started = False
RECONNECT_INTERVAL_S = 15.0
# ── Layer 1 Safety constants ──────────────────────────────────────────────────
OAKD_STOP_DIST = 0.30 # meters — stop if obstacle closer than this
OAKD_SLOW_DIST = 0.60 # meters — slow if obstacle closer than this
FLOOR_CHECK_HZ = 1.0 # floor-drop check rate — 1Hz enough, numpy is heavy
_safety_active = True
_last_floor_check = 0.0
_warmup_frames = 0 # frames received since last (re)connect
WARMUP_FRAMES = 15 # skip void check for first N frames after connect
_cached_front_depth: float | None = None # last valid reading — safe for GUI to poll
# ── Layer 2 YOLO constants ────────────────────────────────────────────────────
YOLO_BLOB_PATH = Path("models/yolov8n_openvino_2022.1_4shave.blob")
YOLO_MIN_CONFIDENCE = 0.55 # minimum detection confidence
YOLO_SLOW_DIST = 3.0 # meters — slow when target within this range
YOLO_STOP_DIST = 2.0 # meters — stop when target within this range
YOLO_CHECK_HZ = 10.0 # detection polling rate
YOLO_COOLDOWN_S = 2.0 # seconds between repeated callbacks, same label
# COCO classes ERIC cares about: people + animals (all large enough to matter)
YOLO_TARGET_CLASSES = {
0: "person",
15: "bird",
16: "cat",
17: "dog",
18: "horse",
19: "sheep",
20: "cow",
21: "elephant",
22: "bear",
23: "zebra",
24: "giraffe",
}
# ── Layer 2 YOLO state ────────────────────────────────────────────────────────
_yolo_spatial_queue = None
_yolo_ok = False
_yolo_callback = None # callable(label, dist_m, bearing, bearing_deg)
_yolo_active = False
_yolo_lock = threading.Lock()
_last_yolo_detect: dict = {} # label → last callback timestamp
_last_yolo_bbox_w: dict = {} # label → last bbox width (approach trend)
_last_yolo_positions: dict = {} # label → (dist_m, bearing, bearing_deg) — latest
_yolo_motor_stop: bool = False # True when Layer 2 issued a motors.stop()
# ── Bearing EMA smoothing ─────────────────────────────────────────────────────
# Single-frame bearing readings are noisy. A 3-frame exponential moving average
# (α=0.4) reduces jitter without adding significant lag.
# Formula: ema = α * new + (1 - α) * prev
_BEARING_EMA_ALPHA = 0.4
_bearing_ema: dict = {} # label → smoothed bearing_deg float
# ── Fatal error strings ───────────────────────────────────────────────────────
_FATAL_ERRORS = (
"MessageQueue was closed",
"X_LINK_ERROR",
"XLinkError",
"Couldn't read data",
"Communication exception",
)
# ─── Public API ───────────────────────────────────────────────────────────────
def oakd_available() -> bool:
return _oakd_ok
def yolo_available() -> bool:
"""True if YOLO pipeline is running on Myriad X."""
return _yolo_ok
def yolo_motor_stop_issued() -> bool:
"""
True if Layer 2 called motors.stop() for a YOLO detection and that
detection has not yet been handled by mission.py.
mission.py calls clear_yolo_motor_stop() after it handles the event
so it knows not to re-issue motors.forward() prematurely.
"""
with _yolo_lock:
return _yolo_motor_stop
def clear_yolo_motor_stop():
"""Call from mission.py after handling a YOLO-triggered stop."""
global _yolo_motor_stop
with _yolo_lock:
_yolo_motor_stop = False
def get_last_yolo_position(label: str) -> dict | None:
"""
Return the most recent YOLO position for a label, regardless of cooldown.
Returns dict with keys: dist_m, bearing, bearing_deg.
Returns None if label has never been detected.
Used by mission.py to recover spatial context after a long Cosmos call.
"""
with _yolo_lock:
pos = _last_yolo_positions.get(label)
return dict(pos) if pos else None
def set_safety_active(active: bool):
"""Enable/disable Layer 1 auto-stop. Use False only for testing."""
global _safety_active
_safety_active = active
log.info(f"OAK-D safety: {'ENABLED' if active else 'DISABLED'}")
def set_yolo_active(active: bool):
"""Enable/disable Layer 2 YOLO detections. Call from mission.py."""
global _yolo_active
_yolo_active = active
log.info(f"OAK-D YOLO detection: {'ACTIVE' if active else 'PAUSED'}")
def set_yolo_callback(fn):
"""
Register mission.py callback for Layer 2 detections.
Signature: fn(label: str, dist_m: float, bearing: str, bearing_deg: float)
bearing — coarse: "left" | "center" | "right"
bearing_deg — precise: signed degrees (−90 … +90), negative=left
Set to None to disable callbacks.
"""
global _yolo_callback
_yolo_callback = fn
log.info(f"OAK-D YOLO callback: {'registered' if fn else 'cleared'}")
# ─── Initialisation ───────────────────────────────────────────────────────────
def init_oakd() -> bool:
"""
(Re-)initialise OAK-D pipeline.
Builds two pipelines on the Myriad X:
1. Stereo depth (Layer 1 safety)
2. YOLOv8n spatial detection (Layer 2 — optional, requires blob)
Safe to call multiple times — tears down existing pipeline first.
"""
global _pipeline, _depth_queue, _oakd_ok, _reader_thread
global _yolo_spatial_queue, _yolo_ok, _warmup_frames
# FIX 4: wait for existing reader thread to exit before rebuilding pipeline
_teardown()
try:
import depthai as dai
_pipeline = dai.Pipeline()
# ── Stereo depth nodes (Layer 1) ──────────────────────────────────────
mono_left = _pipeline.create(dai.node.MonoCamera)
mono_right = _pipeline.create(dai.node.MonoCamera)
stereo = _pipeline.create(dai.node.StereoDepth)
mono_left.setResolution(
dai.MonoCameraProperties.SensorResolution.THE_400_P)
mono_left.setBoardSocket(dai.CameraBoardSocket.LEFT)
mono_left.setFps(8)
mono_right.setResolution(
dai.MonoCameraProperties.SensorResolution.THE_400_P)
mono_right.setBoardSocket(dai.CameraBoardSocket.RIGHT)
mono_right.setFps(8)
stereo.setLeftRightCheck(True)
stereo.setExtendedDisparity(False)
stereo.setSubpixel(False)
# Try best preset first, fall back gracefully
for preset_name in ("HIGH_DENSITY", "HIGH_ACCURACY", "DEFAULT"):
try:
preset = getattr(dai.node.StereoDepth.PresetMode, preset_name)
stereo.setDefaultProfilePreset(preset)
log.info(f"OAK-D stereo preset: {preset_name}")
break
except AttributeError:
continue
mono_left.out.link(stereo.left)
mono_right.out.link(stereo.right)
_depth_queue = stereo.depth.createOutputQueue(maxSize=1, blocking=False)
# ── YOLO spatial detection nodes (Layer 2) ────────────────────────────
_yolo_ok = False
if YOLO_BLOB_PATH.exists():
try:
_yolo_ok = _add_yolo_pipeline(_pipeline, stereo)
if _yolo_ok:
log.info("✅ OAK-D YOLO: YOLOv8n spatial detection active "
"on Myriad X")
except Exception as ye:
log.warning(f"⚠️ OAK-D YOLO pipeline failed ({ye}) "
"— depth still active")
_yolo_ok = False
else:
log.warning(
f"⚠️ OAK-D YOLO: blob not found at {YOLO_BLOB_PATH} "
"— Layer 2 person detection DISABLED. "
"Download from https://github.com/luxonis/depthai-model-zoo"
)
_pipeline.start()
_oakd_ok = True
_warmup_frames = 0 # reset warmup — discard first frames after (re)connect
log.info("✅ OAK-D Lite: stereo depth started (DepthAI 3.x)")
_reader_thread = threading.Thread(
target=_reader_loop, daemon=True, name="oakd-reader"
)
_reader_thread.start()
_init_depth_publisher()
_ensure_reconnect_watchdog()
return True
except Exception as e:
log.warning(f"⚠️ OAK-D init failed ({e}) — depth perception disabled")
_oakd_ok = False
_yolo_ok = False
_ensure_reconnect_watchdog()
return False
def _add_yolo_pipeline(pipeline, stereo) -> bool:
"""
Add YOLOv8n spatial detection to an existing pipeline.
DepthAI 3.x API — YOLO config goes via nn.detectionParser sub-node.
Colour camera feeds YOLO; stereo depth provides spatial coordinates.
FIX 1: ISP scale corrected for OAK-D Lite SIPP memory budget.
The OAK-D Lite IMX214 sensor outputs 2104×1560 raw.
setIspScale(8, 9) targeted ~1870×1386 — not a clean resolution and
still caused E_OUT_OF_MEM + E_FINALISE_FAIL crash loop on the Lite's
lower-spec Myriad X (confirmed in logs: crash every ~5s after reconnect).
setIspScale(1, 2) targets 960×540, which fits comfortably within the
Lite's SIPP buffer limits and leaves headroom for YOLO + stereo depth.
setVideoSize removed — ISP scale drives output dimensions directly.
FIX 5: colour camera FPS reduced to 5 (was raised to 10 in previous fix,
which caused the SIPP OOM on the Lite). The OAK-D Lite Myriad X cannot
sustain 10 FPS colour + YOLO spatial + 15 FPS stereo simultaneously.
5 FPS colour is the validated safe rate for this pipeline on the Lite.
"""
import depthai as dai
cam_rgb = pipeline.create(dai.node.ColorCamera)
cam_rgb.setPreviewSize(320, 320) # Reduced from 416 to ease Myriad X memory
# IMX378/214 only supports 1080_P, 4_K, or 12_MP — other resolutions fall
# back to 1080_P which outputs native 2104×1560. setIspScale ratios on this
# sensor do not land cleanly on standard sizes. Instead: set 1080_P and use
# setVideoSize to crop/scale the ISP output to exactly 960×540, which is
# within the OAK-D Lite SIPP memory budget and what the pipeline expects.
cam_rgb.setResolution(
dai.ColorCameraProperties.SensorResolution.THE_1080_P)
# setVideoSize and setIspScale removed — native 2104x1560 is fine, preview handles YOLO input
cam_rgb.setInterleaved(False)
cam_rgb.setColorOrder(dai.ColorCameraProperties.ColorOrder.BGR)
cam_rgb.setFps(3) # FIX: reduced to 3fps to prevent Myriad X heap corruption
nn = pipeline.create(dai.node.SpatialDetectionNetwork)
nn.setBlobPath(str(YOLO_BLOB_PATH))
nn.setConfidenceThreshold(YOLO_MIN_CONFIDENCE)
nn.setBoundingBoxScaleFactor(0.5)
nn.setDepthLowerThreshold(100) # mm — ignore depth < 10 cm
nn.setDepthUpperThreshold(8000) # mm — ignore depth > 8 m
# DepthAI 3.x — YOLO-specific config lives in detectionParser sub-node
nn.detectionParser.setNumClasses(80)
nn.detectionParser.setCoordinateSize(4)
nn.detectionParser.setAnchors([
10, 13, 16, 30, 33, 23,
30, 61, 62, 45, 59, 119,
116, 90, 156, 198, 373, 326
])
nn.detectionParser.setAnchorMasks({
"side52": [0, 1, 2],
"side26": [3, 4, 5],
"side13": [6, 7, 8]
})
nn.detectionParser.setIouThreshold(0.5)
stereo.depth.link(nn.inputDepth)
cam_rgb.preview.link(nn.input)
global _yolo_spatial_queue
_yolo_spatial_queue = nn.out.createOutputQueue(maxSize=2, blocking=False)
return True
def _teardown():
"""
Stop existing pipeline and clear state without raising.
Joins reader thread (up to 2s) before stopping pipeline.
Note: depth publisher is on ros_core shared node — not torn down here.
ros_core.shutdown() handles it at process exit.
"""
global _pipeline, _depth_queue, _depth_frame, _oakd_ok
global _yolo_spatial_queue, _yolo_ok, _reader_thread
_oakd_ok = False # signal reader loop to exit
_yolo_ok = False
# Wait for reader thread to see _oakd_ok=False and exit cleanly
if _reader_thread is not None and _reader_thread.is_alive():
_reader_thread.join(timeout=2.0)
if _reader_thread.is_alive():
log.warning("OAK-D reader thread did not exit cleanly within 2s")
_reader_thread = None
_depth_queue = None
_yolo_spatial_queue = None
with _lock:
_depth_frame = None
try:
if _pipeline is not None:
_pipeline.stop()
except Exception:
pass
_pipeline = None
# ─── Reader loop ──────────────────────────────────────────────────────────────
def _reader_loop():
"""
Background thread — reads depth frames and YOLO detections continuously.
Layer 1: every depth frame → _safety_check() → instant stop/slow.
Layer 2: every YOLO_CHECK_HZ interval → _yolo_check() → callback.
On fatal disconnect: logs once, clears _oakd_ok, exits cleanly.
Reconnect watchdog handles restart.
"""
global _depth_frame, _oakd_ok, _warmup_frames
yolo_check_interval = 1.0 / YOLO_CHECK_HZ
last_yolo_check = 0.0
while _oakd_ok:
try:
if _depth_queue is None:
time.sleep(0.1)
continue
# 1s timeout — prevents infinite block on disconnect
in_depth = _depth_queue.get(timeout=datetime.timedelta(seconds=1))
if in_depth is None:
continue
frame = in_depth.getFrame()
with _lock:
_depth_frame = frame
_warmup_frames += 1
if _safety_active:
_safety_check(_warmup_frames)
# Publish depth frame for Nav2 costmap (rate-limited)
_maybe_publish_depth(frame)
# ── Layer 2: YOLO detections ──────────────────────────────────────
now = time.monotonic()
if (_yolo_ok and _yolo_active
and _yolo_spatial_queue is not None
and now - last_yolo_check >= yolo_check_interval):
last_yolo_check = now
_yolo_check()
except Exception as e:
err_str = str(e)
if any(token in err_str for token in _FATAL_ERRORS):
log.warning(
f"⚠️ OAK-D disconnected ({err_str.splitlines()[0]}) "
"— depth disabled, reconnect watchdog will retry"
)
_oakd_ok = False
return
log.warning(f"OAK-D reader frame error: {e}")
time.sleep(0.5)
log.info("OAK-D reader loop exited (device offline)")
# ─── Layer 1: Safety check ────────────────────────────────────────────────────
def _safety_check(warmup_frames: int = 999):
"""
Layer 1 safety reactions — called every depth frame.
Check 1 — Front obstacle (3-patch sampling):
Samples depth at (left-of-centre, centre, right-of-centre).
Uses minimum of valid samples — one textureless patch can't blind us.
Logs a warning if fewer than 3 patches return valid depth (possible
glass / white wall) so operators can investigate.
Check 2 — Floor drop / void (rate-limited at FLOOR_CHECK_HZ):
HIGH confidence → motors.stop() — unambiguous drop detected.
MEDIUM confidence is intentionally suppressed here (flat-floor false
positives at 15cm mount height). HIGH only is the safe default.
Both checks are independent of Cosmos — pure reactive hardware safety.
Sensor arbitration:
LiDAR and OAK-D both call motors.stop() independently. Whichever
fires first wins. No arbitration is correct for safety — the most
conservative sensor always dominates.
FIX 2: _avoidance_in_progress is now resolved at the top of the function
so it is always defined before either the obstacle check or the floor-
drop check references it (previously it was set inside the else branch
of the obstacle check, causing UnboundLocalError when valid_patches==0).
"""
global _last_floor_check
# FIX 2: resolve avoidance state once, at the top — always defined
_avoidance_in_progress = False
try:
from lidar import _avoidance_active
_avoidance_in_progress = _avoidance_active
except Exception:
pass
try:
from motors import motors
# ── Check 1: front obstacle — 3-patch depth sampling ─────────────────
samples = _get_front_depth_3patch()
if samples["valid_patches"] == 0:
# No depth at all in front — don't act, log once at debug level
log.debug("OAK-D: 0/3 front depth patches valid — no obstacle action")
else:
if samples["valid_patches"] < 3:
log.debug(
f"OAK-D: only {samples['valid_patches']}/3 depth patches "
"valid (possible glass/white wall) — using available patches"
)
front = samples["min_m"]
if not _avoidance_in_progress:
if front < OAKD_STOP_DIST:
motors.stop()
log.warning(f"🚧 OAK-D STOP — obstacle at {front:.2f}m "
f"({samples['valid_patches']}/3 patches)")
elif front < OAKD_SLOW_DIST:
motors.slow()
log.info(f"⚠️ OAK-D slow — obstacle at {front:.2f}m "
f"({samples['valid_patches']}/3 patches)")
# ── Check 2: floor drop / void — DISABLED for cookoff ───────────────
# False positives on low-texture floors at 15cm mount height.
# LiDAR handles obstacle safety. Re-enable after cookoff with better tuning.
pass
except Exception as e:
log.debug(f"OAK-D safety check error: {e}")
# ─── Layer 2: YOLO check ──────────────────────────────────────────────────────
def _yolo_check():
"""
Layer 2: poll YOLO spatial detections from Myriad X.
Runs at YOLO_CHECK_HZ — only when _yolo_active is True.
For each detection of a target class (person/animal):
1. Extract confidence, spatial depth (mm → m), bounding box.
2. Compute proportional bearing_deg (−45 … +45°) from bbox centre x.
Negative = left, positive = right.
Also compute coarse bearing string (left/center/right) for compat.
3. Track bounding box width trend: if box is SHRINKING (target leaving),
skip callback — don't chase someone walking away.
Only fire callback when box is stable or growing (target approaching
or stationary).
4. Always update _last_yolo_positions regardless of trend or cooldown,
so mission.py can recover spatial context after a long Cosmos call.
5. Layer 2 motor guard: if stopping for a target, set _yolo_motor_stop=True
so mission.py knows not to re-issue motors.forward() immediately.
6. Apply per-label cooldown before firing callback.
"""
global _last_yolo_detect, _last_yolo_bbox_w, _yolo_motor_stop
try:
detections = _yolo_spatial_queue.tryGetAll()
if not detections:
return
# Use most recent detection packet — discard older ones in queue
packet = detections[-1]
now = time.monotonic()
for det in packet.detections:
label_idx = det.label
if label_idx not in YOLO_TARGET_CLASSES:
continue
label = YOLO_TARGET_CLASSES[label_idx]
confidence = det.confidence
if confidence < YOLO_MIN_CONFIDENCE:
continue
# ── Spatial depth ─────────────────────────────────────────────────
dist_mm = det.spatialCoordinates.z
if dist_mm <= 0:
continue
dist_m = dist_mm / 1000.0
# ── Proportional bearing ──────────────────────────────────────────
# bbox centre x: 0.0 = left edge, 1.0 = right edge
# Map to degrees: 0.0→−45°, 0.5→0°, 1.0→+45°
# (±45° is the approximate OAK-D Lite horizontal FOV half-angle)
cx = (det.xmin + det.xmax) / 2.0
bearing_deg = (cx - 0.5) * 90.0 # −45 … +45°, negative=left
if cx < 0.35:
bearing = "left"
elif cx > 0.65:
bearing = "right"
else:
bearing = "center"
# ── Bounding box width trend ──────────────────────────────────────
# bbox width normalised 0.0–1.0
bbox_w = det.xmax - det.xmin
with _yolo_lock:
prev_w = _last_yolo_bbox_w.get(label, bbox_w)
_last_yolo_bbox_w[label] = bbox_w
shrinking = bbox_w < prev_w * 0.85 # >15% smaller → leaving
# ── EMA bearing smoothing ─────────────────────────────────────
# Reduces single-frame jitter without significant lag.
prev_ema = _bearing_ema.get(label, bearing_deg)
smoothed_bearing_deg = (_BEARING_EMA_ALPHA * bearing_deg
+ (1.0 - _BEARING_EMA_ALPHA) * prev_ema)
_bearing_ema[label] = smoothed_bearing_deg
bearing_deg = round(smoothed_bearing_deg, 1)
# Recompute coarse bearing from smoothed value
if bearing_deg < -12.0:
bearing = "left"
elif bearing_deg > 12.0:
bearing = "right"
else:
bearing = "center"
# ── Always update position memory ─────────────────────────────────
with _yolo_lock:
_last_yolo_positions[label] = {
"dist_m": dist_m,
"bearing": bearing,
"bearing_deg": round(bearing_deg, 1),
"confidence": round(confidence, 2),
"timestamp": now,
}
# ── Skip callback if target is clearly moving away ────────────────
if shrinking:
log.debug(f"YOLO: {label} bbox shrinking "
f"({prev_w:.2f}→{bbox_w:.2f}) — target leaving, skip callback")
continue
# ── Layer 2 motor reaction ────────────────────────────────────────
try:
from motors import motors
if dist_m < YOLO_STOP_DIST:
motors.stop()
with _yolo_lock:
_yolo_motor_stop = True
log.info(f"🛑 YOLO STOP — {label} at {dist_m:.1f}m "
f"({bearing} / {bearing_deg:+.0f}°)")
elif dist_m < YOLO_SLOW_DIST:
motors.slow()
log.info(f"⚠️ YOLO slow — {label} at {dist_m:.1f}m "
f"({bearing} / {bearing_deg:+.0f}°)")
except Exception:
pass
# ── Per-label cooldown before callback ────────────────────────────
with _yolo_lock:
last_cb = _last_yolo_detect.get(label, 0.0)
if now - last_cb < YOLO_COOLDOWN_S:
continue
_last_yolo_detect[label] = now
# ── Fire callback ─────────────────────────────────────────────────
cb = _yolo_callback
if cb:
try:
cb(label, dist_m, bearing, bearing_deg)
except Exception as ce:
log.warning(f"YOLO callback error: {ce}")
except Exception as e:
log.debug(f"OAK-D YOLO check error: {e}")
# ─── SLAM / Nav2 depth publishing ────────────────────────────────────────────
def _init_depth_publisher():
"""
Create ROS2 publisher for /oakd/depth on the shared ros_core node.
Nav2 costmap subscribes to this for 3D obstacle awareness.
Fix: previously created its own node ("eric_oakd_depth_pub") or chased
_node imports from other modules. On reconnect, _teardown() never
destroyed the old node, so each reconnect leaked a new node with the
same name. Now uses ros_core.get_node() — one shared node, no leaks,
no duplicate subscriptions.
"""
global _depth_pub
try:
from ros_core import get_node, ensure_spinning
from sensor_msgs.msg import Image
node = get_node()
if node is None:
log.debug("Depth publisher: ROS2 unavailable — skipping")
return
_depth_pub = node.create_publisher(Image, "/oakd/depth", 10)
ensure_spinning()
log.info("📡 OAK-D depth publisher → /oakd/depth (Nav2 costmap)")
except Exception as e:
log.debug(f"Depth publisher init skipped: {e}")
def _maybe_publish_depth(frame):
"""
Publish depth frame to /oakd/depth at _depth_pub_hz rate.
Nav2 costmap layer subscribes to this for obstacle inflation.
Rate-limited to avoid flooding the ROS2 bus.
"""
global _last_depth_pub
if _depth_pub is None:
return
now = time.monotonic()
if now - _last_depth_pub < 1.0 / _depth_pub_hz:
return
_last_depth_pub = now
try:
import numpy as np
from sensor_msgs.msg import Image
from std_msgs.msg import Header
from ros_core import get_node as _get_ros_node
msg = Image()
msg.header.stamp = _get_ros_node().get_clock().now().to_msg()
msg.header.frame_id = "oakd_link"
msg.height = frame.shape[0]
msg.width = frame.shape[1]
msg.encoding = "16UC1" # 16-bit unsigned, millimeters
msg.is_bigendian = False
msg.step = frame.shape[1] * 2
msg.data = frame.astype("uint16").tobytes()
_depth_pub.publish(msg)
except Exception as e:
log.debug(f"Depth publish error: {e}")
# ─── Reconnect watchdog ───────────────────────────────────────────────────────
def _ensure_reconnect_watchdog():
"""
Start the reconnect watchdog thread exactly once.
FIX 3: uses a Lock instead of a bare boolean to prevent two watchdog
threads from spawning if init_oakd() is called concurrently.
"""
global _reconnect_thread, _reconnect_started
with _reconnect_lock:
if _reconnect_started:
return
_reconnect_started = True
_reconnect_thread = threading.Thread(
target=_reconnect_loop, daemon=True, name="oakd-reconnect"
)
_reconnect_thread.start()
def _reconnect_loop():
while True:
time.sleep(RECONNECT_INTERVAL_S)
if not _oakd_ok:
log.info("🔄 OAK-D reconnect attempt...")
init_oakd()
# ─── Depth query helpers ──────────────────────────────────────────────────────
def get_depth_at(x_ratio: float = 0.5, y_ratio: float = 0.5,
patch_px: int = 10) -> float | None:
"""Return median depth (meters) at a normalised (x,y) position."""
if not _oakd_ok:
return None
with _lock:
if _depth_frame is None:
return None
frame = _depth_frame.copy()
h, w = frame.shape
cx = int(x_ratio * w)
cy = int(y_ratio * h)
# FIX 7: +1 on upper bounds so slice width is always 2*patch_px pixels
x1, x2 = max(0, cx - patch_px), min(w, cx + patch_px + 1)
y1, y2 = max(0, cy - patch_px), min(h, cy + patch_px + 1)
patch = frame[y1:y2, x1:x2]
valid = patch[patch > 0]
if len(valid) == 0:
return None
return float(np.median(valid)) / 1000.0
def _get_front_depth_3patch() -> dict:
"""
Sample depth at three horizontal positions in the front-centre zone:
left-of-centre (x=0.35, y=0.65)
centre (x=0.50, y=0.65)
right-of-centre (x=0.65, y=0.65)
Returns dict:
{
"min_m": float — minimum of valid samples (use for safety),
"samples": list[float | None] — [left, centre, right],
"valid_patches": int — number of patches that returned valid depth,
}
Using the minimum of three samples means a single textureless patch
(white wall, glass door) can't silence the whole front obstacle check.
Returns min_m=999.0 if no patches are valid.
"""
positions = [(0.35, 0.65), (0.50, 0.65), (0.65, 0.65)]
samples = [get_depth_at(x, y) for x, y in positions]
valid = [s for s in samples if s is not None]
return {
"min_m": min(valid) if valid else 999.0,
"samples": samples,
"valid_patches": len(valid),
}
def get_front_depth() -> float | None:
"""
Return minimum front depth (meters) from 3-patch sampling.
Caches last valid reading so GUI pollers always get a value even if
called between frames or during warmup.
Returns None only if no valid reading has ever been obtained.
"""
global _cached_front_depth
result = _get_front_depth_3patch()
m = result["min_m"]
if m < 999.0:
_cached_front_depth = m
return m
# No fresh reading — return last known good (stays current enough for display)
return _cached_front_depth
def get_depth_map() -> dict | None:
"""Return a 9-cell depth map (top/mid/bot × left/center/right)."""
if not _oakd_ok:
return None
cells = {
"top_left": (0.2, 0.2), "top_center": (0.5, 0.2), "top_right": (0.8, 0.2),
"mid_left": (0.2, 0.5), "mid_center": (0.5, 0.5), "mid_right": (0.8, 0.5),
"bot_left": (0.2, 0.8), "bot_center": (0.5, 0.8), "bot_right": (0.8, 0.8),
}
result = {}
for name, (x, y) in cells.items():
d = get_depth_at(x, y)
result[name] = round(d, 2) if d is not None else None
return result
# ─── Floor-drop / void detection ─────────────────────────────────────────────
def get_floor_drop(
strip_y_ratio: float = 0.75, # 15cm mount: floor visible at y≈0.65-0.80, not 0.90
strip_height_px: int = 25, # taller strip gives better statistics at this height
patch_cols: int = 5,
normal_max_m: float = 6.0, # larger rooms won't false-trigger
drop_threshold_m: float = 3.0, # 15cm height — small step looks like large drop
) -> dict:
"""
Detect floor discontinuity (holes, stairs, cliffs) from stereo depth.
Camera mount: OAK-D is ~15cm off the ground. At this height the camera
sees the floor at a very shallow angle — the bottom strip of the frame
is floor at 20-40cm. Stereo depth on low-texture carpet/tile at that
angle returns very few valid pixels (specular reflections, poor baseline).
The old 2% sparse threshold fired on every flat indoor floor.
Tuning for 15cm mount:
strip_y_ratio: 0.90 → 0.75 sample higher where floor is more visible
strip_height_px: 15 → 25 more pixels for better statistics
sparse HIGH: 2% → 0.5% only truly empty = open air
drop_threshold: 2.0 → 3.0 shallow angle exaggerates drop apparent size
normal_max_m: 5.0 → 6.0 larger rooms were false-triggering
Confidence levels:
high → return_ratio < 0.5% OR edge depth >> normal_max_m
medium → drop_delta > drop_threshold_m OR no edge returns but mid valid
low → no anomaly
"""
if not _oakd_ok:
return {"void_detected": False, "confidence": "low",
"floor_mid_m": None, "floor_edge_m": None,
"valid_returns": 0, "reason": "OAK-D unavailable"}
with _lock:
if _depth_frame is None:
return {"void_detected": False, "confidence": "low",
"floor_mid_m": None, "floor_edge_m": None,
"valid_returns": 0, "reason": "no depth frame"}
frame = _depth_frame.copy()
h, w = frame.shape
# ── Mid-frame floor reference ─────────────────────────────────────────────
mid_y = int(0.5 * h)
mid_patch = frame[max(0, mid_y - 10):min(h, mid_y + 10), w // 4: 3 * w // 4]
mid_valid = mid_patch[(mid_patch > 0) & (mid_patch < 8000)]
floor_mid_m = (float(np.median(mid_valid)) / 1000.0
if len(mid_valid) > 10 else None)
# ── Bottom-strip floor edge ───────────────────────────────────────────────
edge_y0 = int(strip_y_ratio * h)
edge_y1 = min(h, edge_y0 + strip_height_px)
edge_strip = frame[edge_y0:edge_y1, :]
col_step = w // (patch_cols + 1)
edge_depths = []
valid_count = 0
for col in range(1, patch_cols + 1):
cx = col * col_step
patch = edge_strip[:, max(0, cx - 8):min(w, cx + 8)]
valid = patch[(patch > 0) & (patch < 12000)]
valid_count += len(valid)
if len(valid) >= 3:
edge_depths.append(float(np.median(valid)) / 1000.0)
floor_edge_m = float(np.median(edge_depths)) if edge_depths else None
# ── Void decision ─────────────────────────────────────────────────────────
void_detected = False
confidence = "low"
reason = "no anomaly"
total_edge_pixels = edge_strip.size
return_ratio = valid_count / max(total_edge_pixels, 1)
if return_ratio < 0.001 and total_edge_pixels > 500: # < 0.1% AND large sample — truly open air only
void_detected = True
confidence = "high"
reason = (f"floor edge returns sparse ({return_ratio:.1%}) "
"— drop or hole below")
elif floor_mid_m is not None and floor_edge_m is not None:
drop_delta = floor_edge_m - floor_mid_m
if floor_edge_m > normal_max_m:
void_detected = True
confidence = "high"
reason = (f"floor edge depth {floor_edge_m:.1f}m >> "
f"normal ({floor_mid_m:.1f}m) — stairs or cliff")
elif drop_delta > drop_threshold_m:
void_detected = True
confidence = "medium"
reason = (f"floor drops {drop_delta:.1f}m at edge "
f"({floor_mid_m:.1f}m→{floor_edge_m:.1f}m) "
"— step or slope")
# Disabled: "no edge returns" at 15cm mount height is normal on low-texture floors
# elif floor_edge_m is None and floor_mid_m is not None and floor_mid_m < 2.0:
# void_detected = True
return {
"void_detected": void_detected,
"confidence": confidence,
"floor_mid_m": round(floor_mid_m, 2) if floor_mid_m is not None else None,
"floor_edge_m": round(floor_edge_m, 2) if floor_edge_m is not None else None,
"valid_returns": valid_count,
"reason": reason,
}
def void_ahead() -> bool:
"""Quick boolean — True if any void is detected at any confidence."""
return get_floor_drop()["void_detected"]
# ─── Status helpers ───────────────────────────────────────────────────────────
def get_status() -> dict:
"""Return current OAK-D + YOLO status for GUI or debug display."""
result = _get_front_depth_3patch()
return {
"available": _oakd_ok,