forked from geaxgx/depthai_blazepose
-
Notifications
You must be signed in to change notification settings - Fork 1
/
template_manager_script.py
226 lines (199 loc) · 9.49 KB
/
template_manager_script.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
"""
This file is the template of the scripting node source code in edge mode
Substitution is made in BlazeposeDepthaiEdge.py
In the following:
rrn_ : normalized [0:1] coordinates in rotated rectangle coordinate systems
sqn_ : normalized [0:1] coordinates in squared input image
"""
${_TRACE} ("Starting manager script node")
import marshal
from math import sin, cos, atan2, pi, hypot, degrees, floor
# Indexes of some keypoints
left_shoulder = 11
right_shoulder = 12
left_hip = 23
right_hip = 24
${_IF_XYZ}
# We use a filter for smoothing the reference point (mid hips or mid shoulders)
# for which we fetch the (x,y,z). Without this filter, the reference point is very shaky
class SmoothingFilter:
def __init__(self, alpha):
self.alpha = alpha
self.initialized = False
def apply(self, x):
if self.initialized:
result = self.alpha * x + (1 - self.alpha) * self.prev_x
else:
result = x
self.initialized = True
self.prev_x = result
return int(result)
def reset(self):
self.initialized = False
filter_x = SmoothingFilter(0.5)
filter_y = SmoothingFilter(0.5)
${_IF_XYZ}
# BufferMgr is used to statically allocate buffers once
# (replace dynamic allocation).
# These buffers are used for sending result to host
class BufferMgr:
def __init__(self):
self._bufs = {}
def __call__(self, size):
try:
buf = self._bufs[size]
except KeyError:
buf = self._bufs[size] = Buffer(size)
${_TRACE} (f"New buffer allocated: {size}")
return buf
buffer_mgr = BufferMgr()
def send_result(type, lm_score=0, rect_center_x=0, rect_center_y=0, rect_size=0, rotation=0, lms=0, lms_world=0, xyz_ref=0, xyz=0, xyz_zone=0):
# type : 0, 1 or 2
# 0 : pose detection only (detection score < threshold)
# 1 : pose detection + landmark regression
# 2 : landmark regression only (ROI computed from previous landmarks)
result = dict([("type", type), ("lm_score", lm_score), ("rotation", rotation),
("rect_center_x", rect_center_x), ("rect_center_y", rect_center_y), ("rect_size", rect_size),
("lms", lms), ('lms_world', lms_world),
("xyz_ref", xyz_ref), ("xyz", xyz), ("xyz_zone", xyz_zone)])
result_serial = marshal.dumps(result, 2)
buffer = buffer_mgr(len(result_serial))
buffer.getData()[:] = result_serial
${_TRACE} ("len result:"+str(len(result_serial)))
node.io['host'].send(buffer)
${_TRACE} ("Manager sent result to host")
def rr2img(rrn_x, rrn_y):
# Convert a point (rrn_x, rrn_y) expressed in normalized rotated rectangle (rrn)
# into (X, Y) expressed in normalized image (sqn)
X = sqn_rr_center_x + sqn_rr_size * ((rrn_x - 0.5) * cos_rot + (0.5 - rrn_y) * sin_rot)
Y = sqn_rr_center_y + sqn_rr_size * ((rrn_y - 0.5) * cos_rot + (rrn_x - 0.5) * sin_rot)
return X, Y
norm_pad_size = ${_pad_h} / ${_frame_size}
def is_visible(lm_id):
# Is the landmark lm_id is visible ?
# Here visibility means inferred visibility from the landmark model
return lms[lm_id*5+3] > ${_visibility_threshold}
def is_in_image(sqn_x, sqn_y):
# Is the point (sqn_x, sqn_y) is included in the useful part of the image (excluding the pads)?
return norm_pad_size < sqn_y < 1 - norm_pad_size
# send_new_frame_to_branch defines on which branch new incoming frames are sent
# 1 = pose detection branch
# 2 = landmark branch
send_new_frame_to_branch = 1
next_roi_lm_idx = 33*5
cfg_pre_pd = ImageManipConfig()
cfg_pre_pd.setResizeThumbnail(224, 224, 0, 0, 0)
while True:
if send_new_frame_to_branch == 1: # Routing frame to pd
node.io['pre_pd_manip_cfg'].send(cfg_pre_pd)
${_TRACE} ("Manager sent thumbnail config to pre_pd manip")
# Wait for pd post processing's result
detection = node.io['from_post_pd_nn'].get().getLayerFp16("result")
${_TRACE} ("Manager received pd result: "+str(detection))
pd_score, sqn_rr_center_x, sqn_rr_center_y, sqn_scale_x, sqn_scale_y = detection
if pd_score < ${_pd_score_thresh}:
send_result(0)
continue
scale_center_x = sqn_scale_x - sqn_rr_center_x
scale_center_y = sqn_scale_y - sqn_rr_center_y
sqn_rr_size = 2 * ${_rect_transf_scale} * hypot(scale_center_x, scale_center_y)
rotation = 0.5 * pi - atan2(-scale_center_y, scale_center_x)
rotation = rotation - 2 * pi *floor((rotation + pi) / (2 * pi))
${_IF_XYZ}
filter_x.reset()
filter_y.reset()
${_IF_XYZ}
# Routing frame to lm
sin_rot = sin(rotation) # We will often need these values later
cos_rot = cos(rotation)
# Tell pre_lm_manip how to crop body region
rr = RotatedRect()
rr.center.x = sqn_rr_center_x
rr.center.y = (sqn_rr_center_y * ${_frame_size} - ${_pad_h}) / ${_img_h}
rr.size.width = sqn_rr_size
rr.size.height = sqn_rr_size * ${_frame_size} / ${_img_h}
rr.angle = degrees(rotation)
cfg = ImageManipConfig()
cfg.setCropRotatedRect(rr, True)
cfg.setResize(256, 256)
node.io['pre_lm_manip_cfg'].send(cfg)
${_TRACE} ("Manager sent config to pre_lm manip")
# Wait for lm's result
lm_result = node.io['from_lm_nn'].get()
${_TRACE} ("Manager received result from lm nn")
lm_score = lm_result.getLayerFp16("Identity_1")[0]
if lm_score > ${_lm_score_thresh}:
lms = lm_result.getLayerFp16("Identity")
lms_world = lm_result.getLayerFp16("Identity_4")[:99]
xyz = 0
xyz_zone = 0
xyz_ref = 0
# Query xyz
${_IF_XYZ}
# Choosing the reference point: mid hips if hips visible, or mid shoulders otherwise
# xyz_ref codes the reference point, 1 if mid hips, 2 if mid shoulders, 0 if no reference point
if is_visible(right_hip) and is_visible(left_hip):
kp1 = right_hip
kp2 = left_hip
rrn_xyz_ref_x = (lms[5*kp1] + lms[5*kp2]) / 512 # 512 = 256*2 (256 for normalizing, 2 for the mean)
rrn_xyz_ref_y = (lms[5*kp1+1] + lms[5*kp2+1]) / 512
sqn_xyz_ref_x, sqn_xyz_ref_y = rr2img(rrn_xyz_ref_x, rrn_xyz_ref_y)
if is_in_image(sqn_xyz_ref_x, sqn_xyz_ref_y):
xyz_ref = 1
if xyz_ref == 0 and is_visible(right_shoulder) and is_visible(left_shoulder):
kp1 = right_shoulder
kp2 = left_shoulder
rrn_xyz_ref_x = (lms[5*kp1] + lms[5*kp2]) / 512 # 512 = 256*2 (256 for normalizing, 2 for the mean)
rrn_xyz_ref_y = (lms[5*kp1+1] + lms[5*kp2+1]) / 512
sqn_xyz_ref_x, sqn_xyz_ref_y = rr2img(rrn_xyz_ref_x, rrn_xyz_ref_y)
if is_in_image(sqn_xyz_ref_x, sqn_xyz_ref_y):
xyz_ref = 2
if xyz_ref:
cfg = SpatialLocationCalculatorConfig()
conf_data = SpatialLocationCalculatorConfigData()
conf_data.depthThresholds.lowerThreshold = 100
conf_data.depthThresholds.upperThreshold = 10000
half_zone_size = max(int(sqn_rr_size * ${_frame_size} / 90), 4)
xc = filter_x.apply(sqn_xyz_ref_x * ${_frame_size} + ${_crop_w})
yc = filter_y.apply(sqn_xyz_ref_y * ${_frame_size} - ${_pad_h})
roi_left = max(0, xc - half_zone_size)
roi_right = min(${_img_w}-1, xc + half_zone_size)
roi_top = max(0, yc - half_zone_size)
roi_bottom = min(${_img_h}-1, yc + half_zone_size)
roi_topleft = Point2f(roi_left, roi_top)
roi_bottomright = Point2f(roi_right, roi_bottom)
conf_data.roi = Rect(roi_topleft, roi_bottomright)
cfg = SpatialLocationCalculatorConfig()
cfg.addROI(conf_data)
node.io['spatial_location_config'].send(cfg)
${_TRACE} ("Manager sent ROI to spatial_location_config")
# Wait xyz response
xyz_data = node.io['spatial_data'].get().getSpatialLocations()
${_TRACE} ("Manager received spatial_location")
coords = xyz_data[0].spatialCoordinates
xyz = [float(coords.x), float(coords.y), float(coords.z)]
roi = xyz_data[0].config.roi
xyz_zone = [int(roi.topLeft().x - ${_crop_w}), int(roi.topLeft().y), int(roi.bottomRight().x - ${_crop_w}), int(roi.bottomRight().y)]
else:
xyz = [0.0] * 3
xyz_zone = [0] * 4
${_IF_XYZ}
# Send result to host
send_result(send_new_frame_to_branch, lm_score, sqn_rr_center_x, sqn_rr_center_y, sqn_rr_size, rotation, lms, lms_world, xyz_ref, xyz, xyz_zone)
if not ${_force_detection}:
send_new_frame_to_branch = 2
# Calculate the ROI for next frame
rrn_rr_center_x = lms[next_roi_lm_idx] / 256
rrn_rr_center_y = lms[next_roi_lm_idx+1] / 256
rrn_scale_x = lms[next_roi_lm_idx+5] / 256
rrn_scale_y = lms[next_roi_lm_idx+6] / 256
sqn_scale_x, sqn_scale_y = rr2img(rrn_scale_x, rrn_scale_y)
sqn_rr_center_x, sqn_rr_center_y = rr2img(rrn_rr_center_x, rrn_rr_center_y)
scale_center_x = sqn_scale_x - sqn_rr_center_x
scale_center_y = sqn_scale_y - sqn_rr_center_y
sqn_rr_size = 2 * ${_rect_transf_scale} * hypot(scale_center_x, scale_center_y)
rotation = 0.5 * pi - atan2(-scale_center_y, scale_center_x)
rotation = rotation - 2 * pi *floor((rotation + pi) / (2 * pi))
else:
send_result(send_new_frame_to_branch, lm_score)
send_new_frame_to_branch = 1