def get_root_relative_poses(inference_results):
paf_map, heatmap = inference_results
upsample_ratio = 4
found_poses = extract_poses(heatmap[0:-1], paf_map, upsample_ratio)[0]
poses_2d = []
num_kpt_panoptic = 19
num_kpt = 18
for pose_id in range(found_poses.shape[0]):
if found_poses[pose_id, 3] == -1: # skip pose if does not found neck
continue
pose_2d = np.ones(num_kpt_panoptic * 3 + 1,
dtype=np.float32) * -1 # +1 for pose confidence
for kpt_id in range(num_kpt):
if found_poses[pose_id, kpt_id * 3] != -1:
x_2d, y_2d = found_poses[pose_id, kpt_id * 3:kpt_id * 3 + 2]
conf = found_poses[pose_id, kpt_id * 3 + 2]
pose_2d[map_id_to_panoptic[kpt_id] *
3] = x_2d # just repacking
pose_2d[map_id_to_panoptic[kpt_id] * 3 + 1] = y_2d
pose_2d[map_id_to_panoptic[kpt_id] * 3 + 2] = conf
pose_2d[-1] = found_poses[pose_id, -1]
poses_2d.append(pose_2d)
return np.array(poses_2d)
def get_root_relative_poses(inference_results, upsample_ratio, threshold):
heatmap, paf_map = inference_results
found_poses = extract_poses(heatmap[0:-1], paf_map, upsample_ratio)
# scale coordinates to features space
found_poses[:, 0:-1:3] /= upsample_ratio
found_poses[:, 1:-1:3] /= upsample_ratio
poses_2d = []
num_kpt = 18
for pose_id in range(found_poses.shape[0]):
if found_poses[pose_id, 5] == -1: # skip pose if does not found neck
continue
pose_2d = np.ones(num_kpt * 3 + 1,
dtype=np.float32) * -1 # +1 for pose confidence
for kpt_id in range(num_kpt):
if found_poses[pose_id, kpt_id * 3] != -1:
x_2d, y_2d, conf = found_poses[pose_id,
kpt_id * 3:(kpt_id + 1) * 3]
pose_2d[kpt_id * 3] = x_2d # just repacking
pose_2d[kpt_id * 3 + 1] = y_2d
pose_2d[kpt_id * 3 + 2] = conf
pose_2d[-1] = found_poses[pose_id, -1] # Global confidence
poses_2d.append(pose_2d)
poses_2d = np.array(poses_2d)
return poses_2d
def get_root_relative_poses(inference_results):
features, heatmap, paf_map = inference_results
upsample_ratio = 4
found_poses = extract_poses(heatmap[0:-1], paf_map, upsample_ratio)
# scale coordinates to features space
found_poses[:, 0:-1:3] /= upsample_ratio
found_poses[:, 1:-1:3] /= upsample_ratio
poses_2d = []
num_kpt_panoptic = 19
num_kpt = 18
for pose_id in range(found_poses.shape[0]):
if found_poses[pose_id, 5] == -1: # skip pose if does not found neck
continue
pose_2d = np.ones(num_kpt_panoptic * 3 + 1, dtype=np.float32) * -1 # +1 for pose confidence
for kpt_id in range(num_kpt):
if found_poses[pose_id, kpt_id * 3] != -1:
x_2d, y_2d, conf = found_poses[pose_id, kpt_id * 3:(kpt_id + 1) * 3]
pose_2d[map_id_to_panoptic[kpt_id] * 3] = x_2d # just repacking
pose_2d[map_id_to_panoptic[kpt_id] * 3 + 1] = y_2d
pose_2d[map_id_to_panoptic[kpt_id] * 3 + 2] = conf
pose_2d[-1] = found_poses[pose_id, -1]
poses_2d.append(pose_2d)
poses_2d = np.array(poses_2d)
keypoint_threshold = 0.1
poses_3d = np.ones((len(poses_2d), num_kpt_panoptic * 4), dtype=np.float32) * -1
for pose_id in range(poses_3d.shape[0]):
if poses_2d[pose_id, 2] <= keypoint_threshold:
continue
pose_3d = poses_3d[pose_id]
neck_2d = poses_2d[pose_id, 0:2].astype(np.int32)
# read all pose coordinates at neck location
for kpt_id in range(num_kpt_panoptic):
map_3d = features[kpt_id * 3:(kpt_id + 1) * 3]
pose_3d[kpt_id * 4] = map_3d[0, neck_2d[1], neck_2d[0]]
pose_3d[kpt_id * 4 + 1] = map_3d[1, neck_2d[1], neck_2d[0]]
pose_3d[kpt_id * 4 + 2] = map_3d[2, neck_2d[1], neck_2d[0]]
pose_3d[kpt_id * 4 + 3] = poses_2d[pose_id, kpt_id * 3 + 2]
# refine keypoints coordinates at corresponding limbs locations
for limb in limbs:
for kpt_id_from in limb:
if poses_2d[pose_id, kpt_id_from * 3 + 2] <= keypoint_threshold:
continue
for kpt_id_where in limb:
kpt_from_2d = poses_2d[pose_id, kpt_id_from * 3:kpt_id_from * 3 + 2].astype(np.int32)
map_3d = features[kpt_id_where * 3:(kpt_id_where + 1) * 3]
pose_3d[kpt_id_where * 4] = map_3d[0, kpt_from_2d[1], kpt_from_2d[0]]
pose_3d[kpt_id_where * 4 + 1] = map_3d[1, kpt_from_2d[1], kpt_from_2d[0]]
pose_3d[kpt_id_where * 4 + 2] = map_3d[2, kpt_from_2d[1], kpt_from_2d[0]]
break
poses_3d[:, 0::4] *= AVG_PERSON_HEIGHT
poses_3d[:, 1::4] *= AVG_PERSON_HEIGHT
poses_3d[:, 2::4] *= AVG_PERSON_HEIGHT
return poses_3d, poses_2d
def run(self, ocvimg):
global human_pose_available
if human_pose_available == False:
return None
infres = self.inference(ocvimg)
PAFs = infres['Mconv7_stage2_L1'][0]
HMs = infres['Mconv7_stage2_L2'][0]
people = extract_poses(HMs[:-1], PAFs,
4) # Construct poses from HMs and PAFs
return people
def main():
# Prep for OpenVINO Inference Engine for human pose estimation
ie = IECore()
model_hp = 'intel/human-pose-estimation-0001/FP16/human-pose-estimation-0001'
net_hp = ie.read_network(model=model_hp + '.xml',
weights=model_hp + '.bin')
input_name_hp = next(iter(net_hp.inputs)) # Input blob name "data"
input_shape_hp = net_hp.inputs[input_name_hp].shape # [1,3,256,456]
PAF_blobName = list(net_hp.outputs.keys())[0] # 'Mconv7_stage2_L1'
HM_blobName = list(net_hp.outputs.keys())[1] # 'Mconv7_stage2_L2'
PAF_shape = net_hp.outputs[PAF_blobName].shape # [1,38,32,57]
HM_shape = net_hp.outputs[HM_blobName].shape # [1,19,32,57]
exec_net_hp = ie.load_network(net_hp, 'CPU')
# Open a USB webcam
cam = cv2.VideoCapture(0)
#cam = cv2.VideoCapture('people.264')
if cam.isOpened() == False:
print('Failed to open the input movie file (or a webCam)')
sys.exit(-1)
while cv2.waitKey(1) != 27: # 27 == ESC
ret, img = cam.read()
if ret == False:
return 0
inblob = cv2.resize(
img, (input_shape_hp[3], input_shape_hp[2])) # 3=Width, 2=Height
inblob = inblob.transpose(
(2, 0, 1)) # Change data layout from HWC to CHW
inblob = inblob.reshape(input_shape_hp)
res_hp = exec_net_hp.infer(inputs={input_name_hp:
inblob}) # Infer poses
heatmaps = res_hp[HM_blobName][0]
PAFs = res_hp[PAF_blobName][0]
people = extract_poses(heatmaps[:-1], PAFs,
4) # Construct poses from HMs and PAFs
renderPeople(img, people, 4, 0.2)
cv2.imshow('Result', img)
cv2.destroyAllWindows()
return 0
