-
Notifications
You must be signed in to change notification settings - Fork 79
/
HandTrackerBpf.py
847 lines (760 loc) · 43.7 KB
/
HandTrackerBpf.py
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
import numpy as np
from collections import namedtuple
import mediapipe_utils as mpu
import depthai as dai
import cv2
from pathlib import Path
from FPS import FPS, now
import time
import sys
SCRIPT_DIR = Path(__file__).resolve().parent
PALM_DETECTION_MODEL = str(SCRIPT_DIR / "models/palm_detection_sh4.blob")
LANDMARK_MODEL_FULL = str(SCRIPT_DIR / "models/hand_landmark_full_sh4.blob")
LANDMARK_MODEL_LITE = str(SCRIPT_DIR / "models/hand_landmark_lite_sh4.blob")
LANDMARK_MODEL_SPARSE = str(SCRIPT_DIR / "models/hand_landmark_sparse_sh4.blob")
MOVENET_LIGHTNING_MODEL = str(SCRIPT_DIR / "models/movenet_singlepose_lightning_U8_transpose.blob")
MOVENET_THUNDER_MODEL = str(SCRIPT_DIR / "models/movenet_singlepose_thunder_U8_transpose.blob")
def to_planar(arr: np.ndarray, shape: tuple) -> np.ndarray:
return cv2.resize(arr, shape).transpose(2,0,1)#.flatten()
class HandTrackerBpf:
"""
Mediapipe Hand Tracker for depthai
Arguments:
- input_src: frame source,
- "rgb" or None: OAK* internal color camera,
- "rgb_laconic": same as "rgb" but without sending the frames to the host (Edge mode only),
- a file path of an image or a video,
- an integer (eg 0) for a webcam id,
- pd_model: palm detection model blob file,
- pd_nms_thresh: NMS threshold,
- pd_score: confidence score to determine whether a detection is reliable (a float between 0 and 1).
- use_lm: boolean. When True, run landmark model. Otherwise, only palm detection model is run
- lm_model: landmark model. Either:
- 'full' for LANDMARK_MODEL_FULL,
- 'lite' for LANDMARK_MODEL_LITE,
- 'sparse' for LANDMARK_MODEL_SPARSE,
- a path of a blob file.
- lm_score_thresh : confidence score to determine whether landmarks prediction is reliable (a float between 0 and 1).
- use_world_landmarks: boolean. The landmarks model yields 2 types of 3D coordinates :
- coordinates expressed in pixels in the image, always stored in hand.landmarks,
- coordinates expressed in meters in the world, stored in hand.world_landmarks
only if use_world_landmarks is True.
- solo: boolean, when True detect one hand max (much faster since we run the pose detection model only if no hand was detected in the previous frame)
- xyz : boolean, when True get the (x, y, z) coords of the detected hands (if the device supports depth measure).
- crop : boolean which indicates if square cropping on source images is applied or not
- internal_fps : when using the internal color camera as input source, set its FPS to this value (calling setFps()).
- resolution : sensor resolution "full" (1920x1080) or "ultra" (3840x2160),
- internal_frame_height : when using the internal color camera, set the frame height (calling setIspScale()).
The width is calculated accordingly to height and depends on value of 'crop'
- use_gesture : boolean, when True, recognize hand poses froma predefined set of poses
(ONE, TWO, THREE, FOUR, FIVE, OK, PEACE, FIST)
- body_pre_focusing: "right" or "left" or "group" or "higher". Body pre focusing is the use
of a body pose detector to help to focus on the region of the image that
contains one hand ("left" or "right") or "both" hands.
If not in solo mode, body_pre_focusing is forced to 'group'
- body_model : Movenet single pose model: "lightning", "thunder"
- body_score_thresh : Movenet score thresh
- hands_up_only: boolean. When using body_pre_focusing, if hands_up_only is True, consider only hands for which the wrist keypoint
is above the elbow keypoint.
- single_hand_tolerance_thresh (Duo mode only) : In Duo mode, if there is only one hand in a frame,
in order to know when a second hand will appear you need to run the palm detection
in the following frames. Because palm detection is slow, you may want to delay
the next time you will run it. 'single_hand_tolerance_thresh' is the number of
frames during only one hand is detected before palm detection is run again.
- lm_nb_threads : 1 or 2 (default=2), number of inference threads for the landmark model
- stats : boolean, when True, display some statistics when exiting.
- trace : int, 0 = no trace, otherwise print some debug messages or show output of ImageManip nodes
if trace & 1, print application level info like number of palm detections
"""
def __init__(self, input_src=None,
pd_model=PALM_DETECTION_MODEL,
pd_score_thresh=0.5, pd_nms_thresh=0.3,
use_lm=True,
lm_model="lite",
lm_score_thresh=0.5,
use_world_landmarks=False,
solo=False,
xyz=False,
crop=False,
internal_fps=23,
resolution="full",
internal_frame_height=640,
use_gesture=False,
body_pre_focusing = 'group',
body_model = "thunder",
body_score_thresh=0.2,
hands_up_only=True,
single_hand_tolerance_thresh=10,
lm_nb_threads=2,
stats=False,
trace=0
):
self.pd_model = pd_model
print(f"Palm detection blob : {self.pd_model}")
if use_lm:
if lm_model == "full":
self.lm_model = LANDMARK_MODEL_FULL
elif lm_model == "lite":
self.lm_model = LANDMARK_MODEL_LITE
elif lm_model == "sparse":
self.lm_model = LANDMARK_MODEL_SPARSE
else:
self.lm_model = lm_model
print(f"Landmark blob : {self.lm_model}")
if not use_lm and solo:
print("Warning: solo mode desactivated when not using landmarks")
self.solo = False
else:
self.solo = solo
if self.solo:
print("In Solo mode, # of landmark model threads is forced to 1")
self.lm_nb_threads = 1
self.body_pre_focusing = body_pre_focusing
else:
assert lm_nb_threads in [1, 2]
self.lm_nb_threads = lm_nb_threads
print("In Duo mode, body_pre_focusing is forced to 'group'")
self.body_pre_focusing = "group"
self.body_score_thresh = body_score_thresh
if body_model == "lightning":
self.body_model = MOVENET_LIGHTNING_MODEL
self.body_input_length = 192
else:
self.body_model = MOVENET_THUNDER_MODEL
self.body_input_length = 256
print(f"Body pose blob : {self.body_model}")
self.pd_score_thresh = pd_score_thresh
self.pd_nms_thresh = pd_nms_thresh
self.use_lm = use_lm
self.lm_score_thresh = lm_score_thresh
self.xyz = False
self.crop = crop
self.use_world_landmarks = use_world_landmarks
self.internal_fps = internal_fps
self.stats = stats
self.trace = trace
self.use_gesture = use_gesture
self.single_hand_tolerance_thresh = single_hand_tolerance_thresh
self.device = dai.Device()
if input_src == None or input_src == "rgb" or input_src == "rgb_laconic":
# Note that here (in Host mode), specifying "rgb_laconic" has no effect
# Color camera frames are systematically transferred to the host
self.input_type = "rgb" # OAK* internal color camera
self.internal_fps = internal_fps
print(f"Internal camera FPS set to: {self.internal_fps}")
if resolution == "full":
self.resolution = (1920, 1080)
elif resolution == "ultra":
self.resolution = (3840, 2160)
else:
print(f"Error: {resolution} is not a valid resolution !")
sys.exit()
print("Sensor resolution:", self.resolution)
if xyz:
# Check if the device supports stereo
cameras = self.device.getConnectedCameras()
if dai.CameraBoardSocket.LEFT in cameras and dai.CameraBoardSocket.RIGHT in cameras:
self.xyz = True
else:
print("Warning: depth unavailable on this device, 'xyz' argument is ignored")
self.video_fps = self.internal_fps # Used when saving the output in a video file. Should be close to the real fps
if self.crop:
self.frame_size, self.scale_nd = mpu.find_isp_scale_params(internal_frame_height, self.resolution)
self.img_h = self.img_w = self.frame_size
self.pad_w = self.pad_h = 0
self.crop_w = (int(round(self.resolution[0] * self.scale_nd[0] / self.scale_nd[1])) - self.img_w) // 2
else:
width, self.scale_nd = mpu.find_isp_scale_params(internal_frame_height * self.resolution[0] / self.resolution[1], self.resolution, is_height=False)
self.img_h = int(round(self.resolution[1] * self.scale_nd[0] / self.scale_nd[1]))
self.img_w = int(round(self.resolution[0] * self.scale_nd[0] / self.scale_nd[1]))
self.pad_h = (self.img_w - self.img_h) // 2
self.pad_w = 0
self.frame_size = self.img_w
self.crop_w = 0
print(f"Internal camera image size: {self.img_w} x {self.img_h} - crop_w:{self.crop_w} pad_h: {self.pad_h}")
elif input_src.endswith('.jpg') or input_src.endswith('.png') :
self.input_type= "image"
self.img = cv2.imread(input_src)
self.video_fps = 25
self.img_h, self.img_w = self.img.shape[:2]
else:
self.input_type = "video"
if input_src.isdigit():
input_src = int(input_src)
self.cap = cv2.VideoCapture(input_src)
self.video_fps = int(self.cap.get(cv2.CAP_PROP_FPS))
self.img_w = int(self.cap.get(cv2.CAP_PROP_FRAME_WIDTH))
self.img_h = int(self.cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
print("Video FPS:", self.video_fps)
if self.input_type != "rgb":
print(f"Original frame size: {self.img_w}x{self.img_h}")
if self.crop:
self.frame_size = min(self.img_w, self.img_h)
else:
self.frame_size = max(self.img_w, self.img_h)
self.crop_w = max((self.img_w - self.frame_size) // 2, 0)
if self.crop_w: print("Cropping on width :", self.crop_w)
self.crop_h = max((self.img_h - self.frame_size) // 2, 0)
if self.crop_h: print("Cropping on height :", self.crop_h)
self.pad_w = max((self.frame_size - self.img_w) // 2, 0)
if self.pad_w: print("Padding on width :", self.pad_w)
self.pad_h = max((self.frame_size - self.img_h) // 2, 0)
if self.pad_h: print("Padding on height :", self.pad_h)
if self.crop: self.img_h = self.img_w = self.frame_size
print(f"Frame working size: {self.img_w}x{self.img_h}")
self.bpf = mpu.BodyPreFocusing(
self.img_w, self.img_h,
self.pad_w, self.pad_h,
self.frame_size,
mode = self.body_pre_focusing,
score_thresh=body_score_thresh,
hands_up_only=hands_up_only
)
self.crop_region = self.bpf.init_crop_region
self.previous_handedness = None
self.nb_bpf_inferences = 0
self.glob_bpf_rtrip_time = 0
# Create SSD anchors
self.pd_input_length = 128 # Palm detection
self.anchors = mpu.generate_handtracker_anchors(self.pd_input_length, self.pd_input_length)
self.nb_anchors = self.anchors.shape[0]
print(f"{self.nb_anchors} anchors have been created")
# Define and start pipeline
usb_speed = self.device.getUsbSpeed()
self.device.startPipeline(self.create_pipeline())
print(f"Pipeline started - USB speed: {str(usb_speed).split('.')[-1]}")
# Define data queues
if self.input_type == "rgb":
self.q_video = self.device.getOutputQueue(name="cam_out", maxSize=1, blocking=False)
self.q_bpf_out = self.device.getOutputQueue(name="bpf_out", maxSize=1, blocking=False)
self.q_pd_in = self.device.getInputQueue(name="pd_in")
self.q_pd_out = self.device.getOutputQueue(name="pd_out", maxSize=1, blocking=False)
self.q_manip_cfg = self.device.getInputQueue(name="manip_cfg")
if self.use_lm:
self.q_lm_out = self.device.getOutputQueue(name="lm_out", maxSize=2, blocking=False)
self.q_lm_in = self.device.getInputQueue(name="lm_in")
if self.xyz:
self.q_spatial_data = self.device.getOutputQueue(name="spatial_data_out", maxSize=4, blocking=False)
self.q_spatial_config = self.device.getInputQueue("spatial_calc_config_in")
else:
self.q_bpf_in = self.device.getInputQueue(name="bpf_in")
self.q_bpf_out = self.device.getOutputQueue(name="bpf_out", maxSize=4, blocking=True)
self.q_pd_in = self.device.getInputQueue(name="pd_in")
self.q_pd_out = self.device.getOutputQueue(name="pd_out", maxSize=4, blocking=True)
if self.use_lm:
self.q_lm_out = self.device.getOutputQueue(name="lm_out", maxSize=4, blocking=True)
self.q_lm_in = self.device.getInputQueue(name="lm_in")
self.fps = FPS()
self.nb_frames_pd_inference = 0
self.nb_frames_lm_inference = 0
self.nb_lm_inferences = 0
self.nb_failed_lm_inferences = 0
self.nb_frames_lm_inference_after_landmarks_ROI = 0
self.nb_frames_no_hand = 0
self.nb_spatial_requests = 0
self.glob_pd_rtrip_time = 0
self.glob_lm_rtrip_time = 0
self.glob_spatial_rtrip_time = 0
self.use_previous_landmarks = False
self.hands_from_landmarks = None
self.nb_hands_in_previous_frame = 0
if not self.solo: self.single_hand_count = 0
def create_pipeline(self):
print("Creating pipeline...")
# Start defining a pipeline
pipeline = dai.Pipeline()
pipeline.setOpenVINOVersion(version = dai.OpenVINO.Version.VERSION_2021_4)
if self.input_type == "rgb":
# ColorCamera
print("Creating Color Camera...")
cam = pipeline.createColorCamera()
if self.resolution[0] == 1920:
cam.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
else:
cam.setResolution(dai.ColorCameraProperties.SensorResolution.THE_4_K)
cam.setBoardSocket(dai.CameraBoardSocket.RGB)
cam.setInterleaved(False)
cam.setIspScale(self.scale_nd[0], self.scale_nd[1])
cam.setFps(self.internal_fps)
# Movenet takes RGB input
cam.setColorOrder(dai.ColorCameraProperties.ColorOrder.RGB)
manip = pipeline.createImageManip()
manip.setMaxOutputFrameSize(self.body_input_length*self.body_input_length*3)
manip.setWaitForConfigInput(True)
manip.inputImage.setQueueSize(1)
manip.inputImage.setBlocking(False)
cam.preview.link(manip.inputImage)
if self.crop:
cam.setVideoSize(self.frame_size, self.frame_size)
cam.setPreviewSize(self.frame_size, self.frame_size)
else:
cam.setVideoSize(self.img_w, self.img_h)
cam.setPreviewSize(self.img_w, self.img_h)
manip_cfg_in = pipeline.createXLinkIn()
manip_cfg_in.setStreamName("manip_cfg")
manip_cfg_in.out.link(manip.inputConfig)
cam_out = pipeline.createXLinkOut()
cam_out.setStreamName("cam_out")
cam_out.input.setQueueSize(1)
cam_out.input.setBlocking(False)
cam.video.link(cam_out.input)
if self.xyz:
print("Creating MonoCameras, Stereo and SpatialLocationCalculator nodes...")
# For now, RGB needs fixed focus to properly align with depth.
# The value used during calibration should be used here
calib_data = self.device.readCalibration()
calib_lens_pos = calib_data.getLensPosition(dai.CameraBoardSocket.RGB)
print(f"RGB calibration lens position: {calib_lens_pos}")
cam.initialControl.setManualFocus(calib_lens_pos)
mono_resolution = dai.MonoCameraProperties.SensorResolution.THE_400_P
left = pipeline.createMonoCamera()
left.setBoardSocket(dai.CameraBoardSocket.LEFT)
left.setResolution(mono_resolution)
left.setFps(self.internal_fps)
right = pipeline.createMonoCamera()
right.setBoardSocket(dai.CameraBoardSocket.RIGHT)
right.setResolution(mono_resolution)
right.setFps(self.internal_fps)
stereo = pipeline.createStereoDepth()
stereo.setConfidenceThreshold(230)
# LR-check is required for depth alignment
stereo.setLeftRightCheck(True)
stereo.setDepthAlign(dai.CameraBoardSocket.RGB)
stereo.setSubpixel(False) # subpixel True -> latency
spatial_location_calculator = pipeline.createSpatialLocationCalculator()
spatial_location_calculator.setWaitForConfigInput(True)
spatial_location_calculator.inputDepth.setBlocking(False)
spatial_location_calculator.inputDepth.setQueueSize(1)
spatial_data_out = pipeline.createXLinkOut()
spatial_data_out.setStreamName("spatial_data_out")
spatial_data_out.input.setQueueSize(1)
spatial_data_out.input.setBlocking(False)
spatial_calc_config_in = pipeline.createXLinkIn()
spatial_calc_config_in.setStreamName("spatial_calc_config_in")
left.out.link(stereo.left)
right.out.link(stereo.right)
stereo.depth.link(spatial_location_calculator.inputDepth)
spatial_location_calculator.out.link(spatial_data_out.input)
spatial_calc_config_in.out.link(spatial_location_calculator.inputConfig)
# Create body pose model
print("Creating Body Pose Neural Network...")
bpf_nn = pipeline.createNeuralNetwork()
bpf_nn.setBlobPath(self.body_model)
if self.input_type == "rgb":
bpf_nn.input.setQueueSize(1)
bpf_nn.input.setBlocking(False)
manip.out.link(bpf_nn.input)
else:
bpf_in = pipeline.createXLinkIn()
bpf_in.setStreamName("bpf_in")
bpf_in.out.link(bpf_nn.input)
# Body pose output
bpf_out = pipeline.createXLinkOut()
bpf_out.setStreamName("bpf_out")
bpf_nn.out.link(bpf_out.input)
# Define palm detection model
print("Creating Palm Detection Neural Network...")
pd_nn = pipeline.createNeuralNetwork()
pd_nn.setBlobPath(self.pd_model)
# Increase threads for detection
# pd_nn.setNumInferenceThreads(2)
# Palm detection input
pd_in = pipeline.createXLinkIn()
pd_in.setStreamName("pd_in")
pd_in.out.link(pd_nn.input)
# Palm detection output
pd_out = pipeline.createXLinkOut()
pd_out.setStreamName("pd_out")
pd_nn.out.link(pd_out.input)
# Define hand landmark model
if self.use_lm:
print("Creating Hand Landmark Neural Network...")
lm_nn = pipeline.createNeuralNetwork()
lm_nn.setBlobPath(self.lm_model)
lm_nn.setNumInferenceThreads(self.lm_nb_threads)
# Hand landmark input
self.lm_input_length = 224
lm_in = pipeline.createXLinkIn()
lm_in.setStreamName("lm_in")
lm_in.out.link(lm_nn.input)
# Hand landmark output
lm_out = pipeline.createXLinkOut()
lm_out.setStreamName("lm_out")
lm_nn.out.link(lm_out.input)
print("Pipeline created.")
return pipeline
def bpf_postprocess(self, inference):
kps = np.array(inference.getLayerFp16('Identity')).reshape(-1,3) # 17x3
body = mpu.Body(
scores=kps[:,2],
keypoints_norm=kps[:,[1,0]],
score_thresh=self.body_score_thresh,
crop_region=self.crop_region)
body.next_crop_region = self.bpf.determine_crop_region(body)
focus_zone, hand_zone_label = self.bpf.get_focus_zone(body)
return focus_zone, hand_zone_label, body
def pd_postprocess(self, inference, focus_zone):
scores = np.array(inference.getLayerFp16("classificators"), dtype=np.float16) # 896
bboxes = np.array(inference.getLayerFp16("regressors"), dtype=np.float16).reshape((self.nb_anchors,18)) # 896x18
# Decode bboxes
hands = mpu.decode_bboxes(self.pd_score_thresh, scores, bboxes, self.anchors, best_only=self.solo)
# Non maximum suppression (not needed if solo)
if not self.solo:
hands = mpu.non_max_suppression(hands, self.pd_nms_thresh)
if focus_zone:
zone_size = focus_zone[2] - focus_zone[0]
for h in hands:
box = h.pd_box * zone_size
box[0] += focus_zone[0]
box[1] += focus_zone[1]
h.pd_box = box / self.frame_size
for kp in h.pd_kps:
kp *= zone_size
kp += focus_zone[0:2]
kp /= self.frame_size
if self.use_lm:
mpu.detections_to_rect(hands)
mpu.rect_transformation(hands, self.frame_size, self.frame_size)
return hands
def lm_postprocess(self, hand, inference):
# print(inference.getAllLayerNames())
# The output names of the landmarks model are :
# Identity_1 (1x1) : score
# Identity_2 (1x1) : handedness
# Identity_3 (1x63) : world 3D landmarks (in meters)
# Identity (1x63) : screen 3D landmarks (in pixels)
hand.lm_score = inference.getLayerFp16("Identity_1")[0]
if hand.lm_score > self.lm_score_thresh:
hand.handedness = inference.getLayerFp16("Identity_2")[0]
lm_raw = np.array(inference.getLayerFp16("Identity_dense/BiasAdd/Add")).reshape(-1,3)
# hand.norm_landmarks contains the normalized ([0:1]) 3D coordinates of landmarks in the square rotated body bounding box
hand.norm_landmarks = lm_raw / self.lm_input_length
# hand.norm_landmarks[:,2] /= 0.4
# Now calculate hand.landmarks = the landmarks in the image coordinate system (in pixel)
src = np.array([(0, 0), (1, 0), (1, 1)], dtype=np.float32)
dst = np.array([ (x, y) for x,y in hand.rect_points[1:]], dtype=np.float32) # hand.rect_points[0] is left bottom point and points going clockwise!
mat = cv2.getAffineTransform(src, dst)
lm_xy = np.expand_dims(hand.norm_landmarks[:,:2], axis=0)
# lm_z = hand.norm_landmarks[:,2:3] * hand.rect_w_a / 0.4
hand.landmarks = np.squeeze(cv2.transform(lm_xy, mat)).astype(np.int32)
# World landmarks
if self.use_world_landmarks:
hand.world_landmarks = np.array(inference.getLayerFp16("Identity_3_dense/BiasAdd/Add")).reshape(-1,3)
if self.use_gesture: mpu.recognize_gesture(hand)
def spatial_loc_roi_from_palm_center(self, hand):
half_size = int(hand.pd_box[2] * self.frame_size / 2)
zone_size = max(half_size//2, 8)
rect_center = dai.Point2f(int(hand.pd_box[0]*self.frame_size) + half_size - zone_size//2 + self.crop_w, int(hand.pd_box[1]*self.frame_size) + half_size - zone_size//2 - self.pad_h)
rect_size = dai.Size2f(zone_size, zone_size)
return dai.Rect(rect_center, rect_size)
def spatial_loc_roi_from_wrist_landmark(self, hand):
zone_size = max(int(hand.rect_w_a / 10), 8)
rect_center = dai.Point2f(*(hand.landmarks[0]-np.array((zone_size//2 - self.crop_w, zone_size//2 + self.pad_h))))
rect_size = dai.Size2f(zone_size, zone_size)
return dai.Rect(rect_center, rect_size)
def query_xyz(self, spatial_loc_roi_func):
conf_datas = []
for h in self.hands:
conf_data = dai.SpatialLocationCalculatorConfigData()
conf_data.depthThresholds.lowerThreshold = 100
conf_data.depthThresholds.upperThreshold = 10000
conf_data.roi = spatial_loc_roi_func(h)
conf_datas.append(conf_data)
if len(conf_datas) > 0:
cfg = dai.SpatialLocationCalculatorConfig()
cfg.setROIs(conf_datas)
spatial_rtrip_time = now()
self.q_spatial_config.send(cfg)
# Receives spatial locations
spatial_data = self.q_spatial_data.get().getSpatialLocations()
self.glob_spatial_rtrip_time += now() - spatial_rtrip_time
self.nb_spatial_requests += 1
for i,sd in enumerate(spatial_data):
self.hands[i].xyz_zone = [
int(sd.config.roi.topLeft().x) - self.crop_w,
int(sd.config.roi.topLeft().y),
int(sd.config.roi.bottomRight().x) - self.crop_w,
int(sd.config.roi.bottomRight().y)
]
self.hands[i].xyz = [
sd.spatialCoordinates.x,
sd.spatialCoordinates.y,
sd.spatialCoordinates.z
]
def smart_crop_and_resize(self, frame, crop_region):
"""Crops and resize the image to prepare for the model input."""
cropped = frame[max(0,crop_region.ymin):min(self.img_h,crop_region.ymax), max(0,crop_region.xmin):min(self.img_w,crop_region.xmax)]
if crop_region.xmin < 0 or crop_region.xmax >= self.img_w or crop_region.ymin < 0 or crop_region.ymax >= self.img_h:
# Padding is necessary
cropped = cv2.copyMakeBorder(cropped,
max(0,-crop_region.ymin),
max(0, crop_region.ymax-self.img_h),
max(0,-crop_region.xmin),
max(0, crop_region.xmax-self.img_w),
cv2.BORDER_CONSTANT)
cropped = cv2.resize(cropped, (self.body_input_length, self.body_input_length), interpolation=cv2.INTER_AREA)
return cropped
def next_frame(self):
body = None
hand_zone_label = None
bag = {}
bag["body"] = None
self.fps.update()
focus_zone = None # Used with body pre focusing
if self.input_type == "rgb":
if not self.use_previous_landmarks:
# Send image manip config to the device
# We prepare the input to Movenet = Movenet smart cropping
cfg = dai.ImageManipConfig()
points = [
[self.crop_region.xmin, self.crop_region.ymin],
[self.crop_region.xmax-1, self.crop_region.ymin],
[self.crop_region.xmax-1, self.crop_region.ymax-1],
[self.crop_region.xmin, self.crop_region.ymax-1]]
point2fList = []
for p in points:
pt = dai.Point2f()
pt.x, pt.y = p[0], p[1]
point2fList.append(pt)
cfg.setWarpTransformFourPoints(point2fList, False)
cfg.setResize(self.body_input_length, self.body_input_length)
cfg.setFrameType(dai.ImgFrame.Type.RGB888p)
self.q_manip_cfg.send(cfg)
in_video = self.q_video.get()
video_frame = in_video.getCvFrame()
if self.pad_h:
square_frame = cv2.copyMakeBorder(video_frame, self.pad_h, self.pad_h, self.pad_w, self.pad_w, cv2.BORDER_CONSTANT)
else:
square_frame = video_frame
else:
if self.input_type == "image":
frame = self.img.copy()
else:
ok, frame = self.cap.read()
if not ok:
return None, None, None
# Cropping and/or padding of the video frame
video_frame = frame[self.crop_h:self.crop_h+self.frame_size, self.crop_w:self.crop_w+self.frame_size]
if self.pad_h or self.pad_w:
square_frame = cv2.copyMakeBorder(video_frame, self.pad_h, self.pad_h, self.pad_w, self.pad_w, cv2.BORDER_CONSTANT)
else:
square_frame = video_frame
if not self.use_previous_landmarks:
smart_cropped = self.smart_crop_and_resize(video_frame, self.crop_region)
smart_cropped = cv2.cvtColor(smart_cropped, cv2.COLOR_BGR2RGB)
frame_nn = dai.ImgFrame()
frame_nn.setTimestamp(time.monotonic())
frame_nn.setWidth(self.body_input_length)
frame_nn.setHeight(self.body_input_length)
frame_nn.setData(to_planar(smart_cropped, (self.body_input_length, self.body_input_length)))
self.q_bpf_in.send(frame_nn)
bpf_rtrip_time = now()
if not self.use_previous_landmarks:
# Get body pose
inference = self.q_bpf_out.get()
focus_zone, hand_zone_label, body = self.bpf_postprocess(inference)
if self.trace & 1:
print(f"Body pose - focus zone: {None if focus_zone is None else hand_zone_label}")
self.crop_region = body.next_crop_region
self.nb_bpf_inferences += 1
bag["bpf_inference"] = 1
bag["body"] = body
if focus_zone:
bag["focus_zone"] = focus_zone.copy() # Saved for rendering
# focus_zone is calculated in the original image
# We convert it in the padded square image
focus_zone[0] += self.pad_w
focus_zone[1] += self.pad_h
focus_zone[2] += self.pad_w
focus_zone[3] += self.pad_h
focus_frame = square_frame[focus_zone[1]:focus_zone[3], focus_zone[0]:focus_zone[2]]
if self.trace & 2:
cv2.imshow("BPF frame", focus_frame)
if self.input_type != "rgb":
self.glob_bpf_rtrip_time += now() - bpf_rtrip_time
# Send image to pd_nn
frame_nn = dai.ImgFrame()
frame_nn.setTimestamp(time.monotonic())
frame_nn.setWidth(self.pd_input_length)
frame_nn.setHeight(self.pd_input_length)
frame_nn.setData(to_planar(focus_frame if focus_zone else square_frame, (self.pd_input_length, self.pd_input_length)))
self.q_pd_in.send(frame_nn)
pd_rtrip_time = now()
else:
self.nb_frames_no_hand += 1
return video_frame, [], bag
# Get palm detection
if self.use_previous_landmarks:
self.hands = self.hands_from_landmarks
else:
inference = self.q_pd_out.get()
if self.input_type != "rgb":
self.glob_pd_rtrip_time += now() - pd_rtrip_time
hands = self.pd_postprocess(inference, focus_zone)
if self.trace & 1:
print(f"Palm detection - nb hands detected: {len(hands)}")
self.nb_frames_pd_inference += 1
bag["pd_inference"] = 1
if not self.solo and self.nb_hands_in_previous_frame == 1 and len(hands) <= 1:
self.hands = self.hands_from_landmarks
else:
self.hands = hands
if len(self.hands) == 0: self.nb_frames_no_hand += 1
# Hand landmarks, send requests
if self.use_lm:
nb_lm_inferences = len(self.hands)
if self.use_previous_landmarks: self.nb_frames_lm_inference_after_landmarks_ROI += 1
for i,h in enumerate(self.hands):
img_hand = mpu.warp_rect_img(h.rect_points, square_frame, self.lm_input_length, self.lm_input_length)
nn_data = dai.NNData()
nn_data.setLayer("input_1", to_planar(img_hand, (self.lm_input_length, self.lm_input_length)))
self.q_lm_in.send(nn_data)
if i == 0: lm_rtrip_time = now() # We measure only for the first hand
for i,h in enumerate(self.hands):
inference = self.q_lm_out.get()
if i == 0: self.glob_lm_rtrip_time += now() - lm_rtrip_time
self.lm_postprocess(h, inference)
bag["lm_inference"] = len(self.hands)
self.hands = [ h for h in self.hands if h.lm_score > self.lm_score_thresh]
if self.trace & 1:
print(f"Landmarks - nb hands detected : {len(self.hands)}")
# Check that 2 detected hands do not correspond to the same hand in the image
# That may happen when one hand in the image cross another one
# A simple method is to assure that the center of the rotated rectangles are not too close
if len(self.hands) == 2:
dist_rect_centers = mpu.distance(np.array((self.hands[0].rect_x_center_a, self.hands[0].rect_y_center_a)), np.array((self.hands[1].rect_x_center_a, self.hands[1].rect_y_center_a)))
if dist_rect_centers < 5:
# Keep the hand with higher landmark score
if self.hands[0].lm_score > self.hands[1].lm_score:
self.hands = [self.hands[0]]
else:
self.hands = [self.hands[1]]
if self.trace & 1: print("!!! Removing one hand because too close to the other one")
nb_hands = len(self.hands)
if self.xyz:
self.query_xyz(self.spatial_loc_roi_from_wrist_landmark)
if self.solo:
if nb_hands == 1:
# hand_from_landmarks will be used to initialize the bounding rotated rectangle (ROI) in the next frame
self.hands_from_landmarks = [mpu.hand_landmarks_to_rect(self.hands[0])]
self.use_previous_landmarks = True
# In Body Pre Focusing mode, set handedness from hand_zone_label calculated from bpf_postprocess()
# (much more reliable than inferred handedness from landmark model)
if hand_zone_label is None and self.previous_handedness is not None:
self.hands[0].handedness = self.previous_handedness
elif hand_zone_label in ["right", "left"]:
self.previous_handedness = self.hands[0].handedness = 1 if hand_zone_label == "right" else 0
else:
self.use_previous_landmarks = False
self.previous_handedness = None
else: # Duo
if not self.use_previous_landmarks:
if hand_zone_label is not None:
if nb_hands == 2:
if hand_zone_label == "group": # "group" -> we expect 2 hands
d_h0_right = body.distance_to_wrist(self.hands[0], "right", pad_h=self.pad_h)
d_h1_right = body.distance_to_wrist(self.hands[1], "right", pad_h=self.pad_h)
d_h0_left = body.distance_to_wrist(self.hands[0], "left", pad_h=self.pad_h)
d_h1_left = body.distance_to_wrist(self.hands[1], "left", pad_h=self.pad_h)
if d_h0_left + d_h1_right < d_h0_right + d_h1_left:
self.hands[0].handedness = 0
self.hands[1].handedness = 1
else:
self.hands[0].handedness = 1
self.hands[1].handedness = 0
else: # hand_zone_label == "left" or "right". Normally we should have only 1 hand but just ine case...
d_h0 = body.distance_to_wrist(self.hands[0], hand_zone_label, pad_h=self.pad_h)
d_h1 = body.distance_to_wrist(self.hands[1], hand_zone_label, pad_h=self.pad_h)
if d_h0 < d_h1:
self.hands[0].handedness = 1 if hand_zone_label == "right" else 0
# We arbitrary decide that the other hand has the other handedness
self.hands[1].handedness = 1 - self.hands[0].handedness
else:
self.hands[1].handedness = 1 if hand_zone_label == "right" else 0
# We arbitrary decide that the other hand has the other handedness
self.hands[0].handedness = 1 - self.hands[0].handedness
self.previous_handedness = [self.hands[0].handedness, self.hands[1].handedness]
self.hands_from_landmarks = [mpu.hand_landmarks_to_rect(self.hands[0]), mpu.hand_landmarks_to_rect(self.hands[1])]
self.use_previous_landmarks = True
elif nb_hands == 1:
if hand_zone_label == "group": # We would have expected 2 hands
d_h0_right = body.distance_to_wrist(self.hands[0], "right", pad_h=self.pad_h)
d_h0_left = body.distance_to_wrist(self.hands[0], "left", pad_h=self.pad_h)
self.hands[0].handedness = 1 if d_h0_right < d_h0_left else 0
else: # hand_zone_label == "left" or "right".
self.hands[0].handedness = 1 if hand_zone_label == "right" else 0
self.previous_handedness = [self.hands[0].handedness]
self.hands_from_landmarks = [mpu.hand_landmarks_to_rect(self.hands[0])]
self.use_previous_landmarks = True
else: # nb_hands == 0
self.use_previous_landmarks = False
elif nb_hands != self.nb_hands_in_previous_frame:
# We ask for body detection for the next frame
self.use_previous_landmarks = False
# ...but there is a chance that we don't use body detection result
# if there is only one hand
if nb_hands == 1:
# Actually we have also nb_hands_in_previous_frame = 2
# The current detected hand was detected from one of the 2 hands_from_landmarks
# Which one ? The one which has its handedness (or landmarks or norm_landmarks)
# attribute filled
hand_i = 0 if hasattr(self.hands_from_landmarks[0], "handedness") else 1
self.hands[0].handedness = self.previous_handedness[hand_i]
self.hands_from_landmarks = [mpu.hand_landmarks_to_rect(self.hands[0])]
else:
if nb_hands == 2:
self.hands[0].handedness = self.previous_handedness[0]
self.hands[1].handedness = self.previous_handedness[1]
self.hands_from_landmarks = [mpu.hand_landmarks_to_rect(self.hands[0]), mpu.hand_landmarks_to_rect(self.hands[1])]
# self.use_previous_landmarks = True
elif nb_hands == 1:
self.hands[0].handedness = self.previous_handedness[0]
self.hands_from_landmarks = [mpu.hand_landmarks_to_rect(self.hands[0])]
if self.single_hand_count >= self.single_hand_tolerance_thresh:
self.use_previous_landmarks = False
self.single_hand_count = 0
else:
self.single_hand_count += 1
# self.use_previous_landmarks = True
else:
self.use_previous_landmarks = False
# Stats
if nb_lm_inferences: self.nb_frames_lm_inference += 1
self.nb_lm_inferences += nb_lm_inferences
self.nb_failed_lm_inferences += nb_lm_inferences - nb_hands
for hand in self.hands:
# If we added padding to make the image square, we need to remove this padding from landmark coordinates and from rect_points
if self.pad_h > 0:
hand.landmarks[:,1] -= self.pad_h
for i in range(len(hand.rect_points)):
hand.rect_points[i][1] -= self.pad_h
if self.pad_w > 0:
hand.landmarks[:,0] -= self.pad_w
for i in range(len(hand.rect_points)):
hand.rect_points[i][0] -= self.pad_w
# Set the hand label
hand.label = "right" if hand.handedness > 0.5 else "left"
self.nb_hands_in_previous_frame = nb_hands
else: # not use_lm
if self.xyz:
self.query_xyz(self.spatial_loc_roi_from_palm_center)
return video_frame, self.hands, bag
def exit(self):
self.device.close()
# Print some stats
if self.stats:
nb_frames = self.fps.nb_frames()
print(f"FPS : {self.fps.get_global():.1f} f/s (# frames = {nb_frames})")
print(f"# frames w/ no hand : {self.nb_frames_no_hand} ({100*self.nb_frames_no_hand/nb_frames:.1f}%)")
print(f"# frames w/ palm detection : {self.nb_frames_pd_inference} ({100*self.nb_frames_pd_inference/nb_frames:.1f}%)")
print(f"# frames w/ landmark inference : {self.nb_frames_lm_inference} ({100*self.nb_frames_lm_inference/nb_frames:.1f}%)- # after palm detection: {self.nb_frames_lm_inference - self.nb_frames_lm_inference_after_landmarks_ROI} - # after landmarks ROI prediction: {self.nb_frames_lm_inference_after_landmarks_ROI}")
if not self.solo:
print(f"On frames with at least one landmark inference, average number of landmarks inferences/frame: {self.nb_lm_inferences/self.nb_frames_lm_inference:.2f}")
print(f"# lm inferences: {self.nb_lm_inferences} - # failed lm inferences: {self.nb_failed_lm_inferences} ({100*self.nb_failed_lm_inferences/self.nb_lm_inferences:.1f}%)")
if self.input_type != "rgb":
print(f"Body pose estimation round trip : {self.glob_bpf_rtrip_time/self.nb_bpf_inferences*1000:.1f} ms")
print(f"Palm detection round trip : {self.glob_pd_rtrip_time/self.nb_frames_pd_inference*1000:.1f} ms")
if self.use_lm and self.nb_lm_inferences:
print(f"Hand landmark round trip : {self.glob_lm_rtrip_time/self.nb_frames_lm_inference*1000:.1f} ms")
if self.xyz:
print(f"Spatial location requests round trip : {self.glob_spatial_rtrip_time/self.nb_spatial_requests*1000:.1f} ms")