Live Human Pose Estimation with OpenVINO™#
This Jupyter notebook can be launched on-line, opening an interactive environment in a browser window. You can also make a local installation. Choose one of the following options:
This notebook demonstrates live pose estimation with OpenVINO, using the OpenPose human-pose-estimation-0001 model from Open Model Zoo. Final part of this notebook shows live inference results from a webcam. Additionally, you can also upload a video file.
NOTE: To use a webcam, you must run this Jupyter notebook on a computer with a webcam. If you run on a server, the webcam will not work. However, you can still do inference on a video in the final step.
Table of contents:
Installation Instructions#
This is a self-contained example that relies solely on its own code.
We recommend running the notebook in a virtual environment. You only need a Jupyter server to start. For details, please refer to Installation Guide.
%pip install -q "openvino>=2023.1.0" opencv-python tqdm
Note: you may need to restart the kernel to use updated packages.
Imports#
import collections
import time
from pathlib import Path
import cv2
import numpy as np
from IPython import display
from numpy.lib.stride_tricks import as_strided
import openvino as ov
# Fetch `notebook_utils` module
import requests
r = requests.get(
url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/notebook_utils.py",
)
open("notebook_utils.py", "w").write(r.text)
import notebook_utils as utils
The model#
Download the model#
Use the download_file, a function from the notebook_utils file.
It automatically creates a directory structure and downloads the
selected model.
If you want to download another model, replace the name of the model and precision in the code below.
NOTE: This may require a different pose decoder.
# A directory where the model will be downloaded.
base_model_dir = Path("model")
# The name of the model from Open Model Zoo.
model_name = "human-pose-estimation-0001"
# Selected precision (FP32, FP16, FP16-INT8).
precision = "FP16-INT8"
model_path = base_model_dir / "intel" / model_name / precision / f"{model_name}.xml"
if not model_path.exists():
model_url_dir = f"https://storage.openvinotoolkit.org/repositories/open_model_zoo/2022.1/models_bin/3/{model_name}/{precision}/"
utils.download_file(model_url_dir + model_name + ".xml", model_path.name, model_path.parent)
utils.download_file(
model_url_dir + model_name + ".bin",
model_path.with_suffix(".bin").name,
model_path.parent,
)
model/intel/human-pose-estimation-0001/FP16-INT8/human-pose-estimation-0001.xml: 0%| | 0.00/474k [0…
model/intel/human-pose-estimation-0001/FP16-INT8/human-pose-estimation-0001.bin: 0%| | 0.00/4.03M […
Load the model#
Downloaded models are located in a fixed structure, which indicates a vendor, the name of the model and a precision.
Only a few lines of code are required to run the model. First,
initialize OpenVINO Runtime. Then, read the network architecture and
model weights from the .bin and .xml files to compile it for the
desired device. Select device from dropdown list for running inference
using OpenVINO.
device = utils.device_widget()
device
Dropdown(description='Device:', index=1, options=('CPU', 'AUTO'), value='AUTO')
import openvino.properties.hint as hints
# Initialize OpenVINO Runtime
core = ov.Core()
# Read the network from a file.
model = core.read_model(model_path)
# Let the AUTO device decide where to load the model (you can use CPU, GPU as well).
compiled_model = core.compile_model(model=model, device_name=device.value, config={hints.performance_mode(): hints.PerformanceMode.LATENCY})
# Get the input and output names of nodes.
input_layer = compiled_model.input(0)
output_layers = compiled_model.outputs
# Get the input size.
height, width = list(input_layer.shape)[2:]
Input layer has the name of the input node and output layers contain names of output nodes of the network. In the case of OpenPose Model, there is 1 input and 2 outputs: PAFs and keypoints heatmap.
input_layer.any_name, [o.any_name for o in output_layers]
('data', ['Mconv7_stage2_L1', 'Mconv7_stage2_L2'])
OpenPose Decoder#
To transform the raw results from the neural network into pose
estimations, you need OpenPose Decoder. It is provided in the Open
Model
Zoo
and compatible with the human-pose-estimation-0001 model.
If you choose a model other than human-pose-estimation-0001 you will
need another decoder (for example, AssociativeEmbeddingDecoder),
which is available in the demos
section
of Open Model Zoo.
# code from https://github.com/openvinotoolkit/open_model_zoo/blob/9296a3712069e688fe64ea02367466122c8e8a3b/demos/common/python/models/open_pose.py#L135
class OpenPoseDecoder:
BODY_PARTS_KPT_IDS = (
(1, 2),
(1, 5),
(2, 3),
(3, 4),
(5, 6),
(6, 7),
(1, 8),
(8, 9),
(9, 10),
(1, 11),
(11, 12),
(12, 13),
(1, 0),
(0, 14),
(14, 16),
(0, 15),
(15, 17),
(2, 16),
(5, 17),
)
BODY_PARTS_PAF_IDS = (
12,
20,
14,
16,
22,
24,
0,
2,
4,
6,
8,
10,
28,
30,
34,
32,
36,
18,
26,
)
def __init__(
self,
num_joints=18,
skeleton=BODY_PARTS_KPT_IDS,
paf_indices=BODY_PARTS_PAF_IDS,
max_points=100,
score_threshold=0.1,
min_paf_alignment_score=0.05,
delta=0.5,
):
self.num_joints = num_joints
self.skeleton = skeleton
self.paf_indices = paf_indices
self.max_points = max_points
self.score_threshold = score_threshold
self.min_paf_alignment_score = min_paf_alignment_score
self.delta = delta
self.points_per_limb = 10
self.grid = np.arange(self.points_per_limb, dtype=np.float32).reshape(1, -1, 1)
def __call__(self, heatmaps, nms_heatmaps, pafs):
batch_size, _, h, w = heatmaps.shape
assert batch_size == 1, "Batch size of 1 only supported"
keypoints = self.extract_points(heatmaps, nms_heatmaps)
pafs = np.transpose(pafs, (0, 2, 3, 1))
if self.delta > 0:
for kpts in keypoints:
kpts[:, :2] += self.delta
np.clip(kpts[:, 0], 0, w - 1, out=kpts[:, 0])
np.clip(kpts[:, 1], 0, h - 1, out=kpts[:, 1])
pose_entries, keypoints = self.group_keypoints(keypoints, pafs, pose_entry_size=self.num_joints + 2)
poses, scores = self.convert_to_coco_format(pose_entries, keypoints)
if len(poses) > 0:
poses = np.asarray(poses, dtype=np.float32)
poses = poses.reshape((poses.shape[0], -1, 3))
else:
poses = np.empty((0, 17, 3), dtype=np.float32)
scores = np.empty(0, dtype=np.float32)
return poses, scores
def extract_points(self, heatmaps, nms_heatmaps):
batch_size, channels_num, h, w = heatmaps.shape
assert batch_size == 1, "Batch size of 1 only supported"
assert channels_num >= self.num_joints
xs, ys, scores = self.top_k(nms_heatmaps)
masks = scores > self.score_threshold
all_keypoints = []
keypoint_id = 0
for k in range(self.num_joints):
# Filter low-score points.
mask = masks[0, k]
x = xs[0, k][mask].ravel()
y = ys[0, k][mask].ravel()
score = scores[0, k][mask].ravel()
n = len(x)
if n == 0:
all_keypoints.append(np.empty((0, 4), dtype=np.float32))
continue
# Apply quarter offset to improve localization accuracy.
x, y = self.refine(heatmaps[0, k], x, y)
np.clip(x, 0, w - 1, out=x)
np.clip(y, 0, h - 1, out=y)
# Pack resulting points.
keypoints = np.empty((n, 4), dtype=np.float32)
keypoints[:, 0] = x
keypoints[:, 1] = y
keypoints[:, 2] = score
keypoints[:, 3] = np.arange(keypoint_id, keypoint_id + n)
keypoint_id += n
all_keypoints.append(keypoints)
return all_keypoints
def top_k(self, heatmaps):
N, K, _, W = heatmaps.shape
heatmaps = heatmaps.reshape(N, K, -1)
# Get positions with top scores.
ind = heatmaps.argpartition(-self.max_points, axis=2)[:, :, -self.max_points :]
scores = np.take_along_axis(heatmaps, ind, axis=2)
# Keep top scores sorted.
subind = np.argsort(-scores, axis=2)
ind = np.take_along_axis(ind, subind, axis=2)
scores = np.take_along_axis(scores, subind, axis=2)
y, x = np.divmod(ind, W)
return x, y, scores
@staticmethod
def refine(heatmap, x, y):
h, w = heatmap.shape[-2:]
valid = np.logical_and(np.logical_and(x > 0, x < w - 1), np.logical_and(y > 0, y < h - 1))
xx = x[valid]
yy = y[valid]
dx = np.sign(heatmap[yy, xx + 1] - heatmap[yy, xx - 1], dtype=np.float32) * 0.25
dy = np.sign(heatmap[yy + 1, xx] - heatmap[yy - 1, xx], dtype=np.float32) * 0.25
x = x.astype(np.float32)
y = y.astype(np.float32)
x[valid] += dx
y[valid] += dy
return x, y
@staticmethod
def is_disjoint(pose_a, pose_b):
pose_a = pose_a[:-2]
pose_b = pose_b[:-2]
return np.all(np.logical_or.reduce((pose_a == pose_b, pose_a < 0, pose_b < 0)))
def update_poses(
self,
kpt_a_id,
kpt_b_id,
all_keypoints,
connections,
pose_entries,
pose_entry_size,
):
for connection in connections:
pose_a_idx = -1
pose_b_idx = -1
for j, pose in enumerate(pose_entries):
if pose[kpt_a_id] == connection[0]:
pose_a_idx = j
if pose[kpt_b_id] == connection[1]:
pose_b_idx = j
if pose_a_idx < 0 and pose_b_idx < 0:
# Create new pose entry.
pose_entry = np.full(pose_entry_size, -1, dtype=np.float32)
pose_entry[kpt_a_id] = connection[0]
pose_entry[kpt_b_id] = connection[1]
pose_entry[-1] = 2
pose_entry[-2] = np.sum(all_keypoints[connection[0:2], 2]) + connection[2]
pose_entries.append(pose_entry)
elif pose_a_idx >= 0 and pose_b_idx >= 0 and pose_a_idx != pose_b_idx:
# Merge two poses are disjoint merge them, otherwise ignore connection.
pose_a = pose_entries[pose_a_idx]
pose_b = pose_entries[pose_b_idx]
if self.is_disjoint(pose_a, pose_b):
pose_a += pose_b
pose_a[:-2] += 1
pose_a[-2] += connection[2]
del pose_entries[pose_b_idx]
elif pose_a_idx >= 0 and pose_b_idx >= 0:
# Adjust score of a pose.
pose_entries[pose_a_idx][-2] += connection[2]
elif pose_a_idx >= 0:
# Add a new limb into pose.
pose = pose_entries[pose_a_idx]
if pose[kpt_b_id] < 0:
pose[-2] += all_keypoints[connection[1], 2]
pose[kpt_b_id] = connection[1]
pose[-2] += connection[2]
pose[-1] += 1
elif pose_b_idx >= 0:
# Add a new limb into pose.
pose = pose_entries[pose_b_idx]
if pose[kpt_a_id] < 0:
pose[-2] += all_keypoints[connection[0], 2]
pose[kpt_a_id] = connection[0]
pose[-2] += connection[2]
pose[-1] += 1
return pose_entries
@staticmethod
def connections_nms(a_idx, b_idx, affinity_scores):
# From all retrieved connections that share starting/ending keypoints leave only the top-scoring ones.
order = affinity_scores.argsort()[::-1]
affinity_scores = affinity_scores[order]
a_idx = a_idx[order]
b_idx = b_idx[order]
idx = []
has_kpt_a = set()
has_kpt_b = set()
for t, (i, j) in enumerate(zip(a_idx, b_idx)):
if i not in has_kpt_a and j not in has_kpt_b:
idx.append(t)
has_kpt_a.add(i)
has_kpt_b.add(j)
idx = np.asarray(idx, dtype=np.int32)
return a_idx[idx], b_idx[idx], affinity_scores[idx]
def group_keypoints(self, all_keypoints_by_type, pafs, pose_entry_size=20):
all_keypoints = np.concatenate(all_keypoints_by_type, axis=0)
pose_entries = []
# For every limb.
for part_id, paf_channel in enumerate(self.paf_indices):
kpt_a_id, kpt_b_id = self.skeleton[part_id]
kpts_a = all_keypoints_by_type[kpt_a_id]
kpts_b = all_keypoints_by_type[kpt_b_id]
n = len(kpts_a)
m = len(kpts_b)
if n == 0 or m == 0:
continue
# Get vectors between all pairs of keypoints, i.e. candidate limb vectors.
a = kpts_a[:, :2]
a = np.broadcast_to(a[None], (m, n, 2))
b = kpts_b[:, :2]
vec_raw = (b[:, None, :] - a).reshape(-1, 1, 2)
# Sample points along every candidate limb vector.
steps = 1 / (self.points_per_limb - 1) * vec_raw
points = steps * self.grid + a.reshape(-1, 1, 2)
points = points.round().astype(dtype=np.int32)
x = points[..., 0].ravel()
y = points[..., 1].ravel()
# Compute affinity score between candidate limb vectors and part affinity field.
part_pafs = pafs[0, :, :, paf_channel : paf_channel + 2]
field = part_pafs[y, x].reshape(-1, self.points_per_limb, 2)
vec_norm = np.linalg.norm(vec_raw, ord=2, axis=-1, keepdims=True)
vec = vec_raw / (vec_norm + 1e-6)
affinity_scores = (field * vec).sum(-1).reshape(-1, self.points_per_limb)
valid_affinity_scores = affinity_scores > self.min_paf_alignment_score
valid_num = valid_affinity_scores.sum(1)
affinity_scores = (affinity_scores * valid_affinity_scores).sum(1) / (valid_num + 1e-6)
success_ratio = valid_num / self.points_per_limb
# Get a list of limbs according to the obtained affinity score.
valid_limbs = np.where(np.logical_and(affinity_scores > 0, success_ratio > 0.8))[0]
if len(valid_limbs) == 0:
continue
b_idx, a_idx = np.divmod(valid_limbs, n)
affinity_scores = affinity_scores[valid_limbs]
# Suppress incompatible connections.
a_idx, b_idx, affinity_scores = self.connections_nms(a_idx, b_idx, affinity_scores)
connections = list(
zip(
kpts_a[a_idx, 3].astype(np.int32),
kpts_b[b_idx, 3].astype(np.int32),
affinity_scores,
)
)
if len(connections) == 0:
continue
# Update poses with new connections.
pose_entries = self.update_poses(
kpt_a_id,
kpt_b_id,
all_keypoints,
connections,
pose_entries,
pose_entry_size,
)
# Remove poses with not enough points.
pose_entries = np.asarray(pose_entries, dtype=np.float32).reshape(-1, pose_entry_size)
pose_entries = pose_entries[pose_entries[:, -1] >= 3]
return pose_entries, all_keypoints
@staticmethod
def convert_to_coco_format(pose_entries, all_keypoints):
num_joints = 17
coco_keypoints = []
scores = []
for pose in pose_entries:
if len(pose) == 0:
continue
keypoints = np.zeros(num_joints * 3)
reorder_map = [0, -1, 6, 8, 10, 5, 7, 9, 12, 14, 16, 11, 13, 15, 2, 1, 4, 3]
person_score = pose[-2]
for keypoint_id, target_id in zip(pose[:-2], reorder_map):
if target_id < 0:
continue
cx, cy, score = 0, 0, 0 # keypoint not found
if keypoint_id != -1:
cx, cy, score = all_keypoints[int(keypoint_id), 0:3]
keypoints[target_id * 3 + 0] = cx
keypoints[target_id * 3 + 1] = cy
keypoints[target_id * 3 + 2] = score
coco_keypoints.append(keypoints)
scores.append(person_score * max(0, (pose[-1] - 1))) # -1 for 'neck'
return np.asarray(coco_keypoints), np.asarray(scores)
Processing#
decoder = OpenPoseDecoder()
Process Results#
A bunch of useful functions to transform results into poses.
First, pool the heatmap. Since pooling is not available in numpy, use a simple method to do it directly with numpy. Then, use non-maximum suppression to get the keypoints from the heatmap. After that, decode poses by using the decoder. Since the input image is bigger than the network outputs, you need to multiply all pose coordinates by a scaling factor.
# 2D pooling in numpy (from: https://stackoverflow.com/a/54966908/1624463)
def pool2d(A, kernel_size, stride, padding, pool_mode="max"):
"""
2D Pooling
Parameters:
A: input 2D array
kernel_size: int, the size of the window
stride: int, the stride of the window
padding: int, implicit zero paddings on both sides of the input
pool_mode: string, 'max' or 'avg'
"""
# Padding
A = np.pad(A, padding, mode="constant")
# Window view of A
output_shape = (
(A.shape[0] - kernel_size) // stride + 1,
(A.shape[1] - kernel_size) // stride + 1,
)
kernel_size = (kernel_size, kernel_size)
A_w = as_strided(
A,
shape=output_shape + kernel_size,
strides=(stride * A.strides[0], stride * A.strides[1]) + A.strides,
)
A_w = A_w.reshape(-1, *kernel_size)
# Return the result of pooling.
if pool_mode == "max":
return A_w.max(axis=(1, 2)).reshape(output_shape)
elif pool_mode == "avg":
return A_w.mean(axis=(1, 2)).reshape(output_shape)
# non maximum suppression
def heatmap_nms(heatmaps, pooled_heatmaps):
return heatmaps * (heatmaps == pooled_heatmaps)
# Get poses from results.
def process_results(img, pafs, heatmaps):
# This processing comes from
# https://github.com/openvinotoolkit/open_model_zoo/blob/master/demos/common/python/models/open_pose.py
pooled_heatmaps = np.array([[pool2d(h, kernel_size=3, stride=1, padding=1, pool_mode="max") for h in heatmaps[0]]])
nms_heatmaps = heatmap_nms(heatmaps, pooled_heatmaps)
# Decode poses.
poses, scores = decoder(heatmaps, nms_heatmaps, pafs)
output_shape = list(compiled_model.output(index=0).partial_shape)
output_scale = (
img.shape[1] / output_shape[3].get_length(),
img.shape[0] / output_shape[2].get_length(),
)
# Multiply coordinates by a scaling factor.
poses[:, :, :2] *= output_scale
return poses, scores
Draw Pose Overlays#
Draw pose overlays on the image to visualize estimated poses. Joints are drawn as circles and limbs are drawn as lines. The code is based on the Human Pose Estimation Demo from Open Model Zoo.
colors = (
(255, 0, 0),
(255, 0, 255),
(170, 0, 255),
(255, 0, 85),
(255, 0, 170),
(85, 255, 0),
(255, 170, 0),
(0, 255, 0),
(255, 255, 0),
(0, 255, 85),
(170, 255, 0),
(0, 85, 255),
(0, 255, 170),
(0, 0, 255),
(0, 255, 255),
(85, 0, 255),
(0, 170, 255),
)
default_skeleton = (
(15, 13),
(13, 11),
(16, 14),
(14, 12),
(11, 12),
(5, 11),
(6, 12),
(5, 6),
(5, 7),
(6, 8),
(7, 9),
(8, 10),
(1, 2),
(0, 1),
(0, 2),
(1, 3),
(2, 4),
(3, 5),
(4, 6),
)
def draw_poses(img, poses, point_score_threshold, skeleton=default_skeleton):
if poses.size == 0:
return img
img_limbs = np.copy(img)
for pose in poses:
points = pose[:, :2].astype(np.int32)
points_scores = pose[:, 2]
# Draw joints.
for i, (p, v) in enumerate(zip(points, points_scores)):
if v > point_score_threshold:
cv2.circle(img, tuple(p), 1, colors[i], 2)
# Draw limbs.
for i, j in skeleton:
if points_scores[i] > point_score_threshold and points_scores[j] > point_score_threshold:
cv2.line(
img_limbs,
tuple(points[i]),
tuple(points[j]),
color=colors[j],
thickness=4,
)
cv2.addWeighted(img, 0.4, img_limbs, 0.6, 0, dst=img)
return img
Main Processing Function#
Run pose estimation on the specified source. Either a webcam or a video file.
# Main processing function to run pose estimation.
def run_pose_estimation(source=0, flip=False, use_popup=False, skip_first_frames=0):
pafs_output_key = compiled_model.output("Mconv7_stage2_L1")
heatmaps_output_key = compiled_model.output("Mconv7_stage2_L2")
player = None
try:
# Create a video player to play with target fps.
player = utils.VideoPlayer(source, flip=flip, fps=30, skip_first_frames=skip_first_frames)
# Start capturing.
player.start()
if use_popup:
title = "Press ESC to Exit"
cv2.namedWindow(title, cv2.WINDOW_GUI_NORMAL | cv2.WINDOW_AUTOSIZE)
processing_times = collections.deque()
while True:
# Grab the frame.
frame = player.next()
if frame is None:
print("Source ended")
break
# If the frame is larger than full HD, reduce size to improve the performance.
scale = 1280 / max(frame.shape)
if scale < 1:
frame = cv2.resize(frame, None, fx=scale, fy=scale, interpolation=cv2.INTER_AREA)
# Resize the image and change dims to fit neural network input.
# (see https://github.com/openvinotoolkit/open_model_zoo/tree/master/models/intel/human-pose-estimation-0001)
input_img = cv2.resize(frame, (width, height), interpolation=cv2.INTER_AREA)
# Create a batch of images (size = 1).
input_img = input_img.transpose((2, 0, 1))[np.newaxis, ...]
# Measure processing time.
start_time = time.time()
# Get results.
results = compiled_model([input_img])
stop_time = time.time()
pafs = results[pafs_output_key]
heatmaps = results[heatmaps_output_key]
# Get poses from network results.
poses, scores = process_results(frame, pafs, heatmaps)
# Draw poses on a frame.
frame = draw_poses(frame, poses, 0.1)
processing_times.append(stop_time - start_time)
# Use processing times from last 200 frames.
if len(processing_times) > 200:
processing_times.popleft()
_, f_width = frame.shape[:2]
# mean processing time [ms]
processing_time = np.mean(processing_times) * 1000
fps = 1000 / processing_time
cv2.putText(
frame,
f"Inference time: {processing_time:.1f}ms ({fps:.1f} FPS)",
(20, 40),
cv2.FONT_HERSHEY_COMPLEX,
f_width / 1000,
(0, 0, 255),
1,
cv2.LINE_AA,
)
# Use this workaround if there is flickering.
if use_popup:
cv2.imshow(title, frame)
key = cv2.waitKey(1)
# escape = 27
if key == 27:
break
else:
# Encode numpy array to jpg.
_, encoded_img = cv2.imencode(".jpg", frame, params=[cv2.IMWRITE_JPEG_QUALITY, 90])
# Create an IPython image.
i = display.Image(data=encoded_img)
# Display the image in this notebook.
display.clear_output(wait=True)
display.display(i)
# ctrl-c
except KeyboardInterrupt:
print("Interrupted")
# any different error
except RuntimeError as e:
print(e)
finally:
if player is not None:
# Stop capturing.
player.stop()
if use_popup:
cv2.destroyAllWindows()
Run#
Run Live Pose Estimation#
Use a webcam as the video input. By default, the primary webcam is set
with source=0. If you have multiple webcams, each one will be
assigned a consecutive number starting at 0. Set flip=True when
using a front-facing camera. Some web browsers, especially Mozilla
Firefox, may cause flickering. If you experience flickering, set
use_popup=True.
NOTE: To use this notebook with a webcam, you need to run the notebook on a computer with a webcam. If you run the notebook on a server (for example, Binder), the webcam will not work. Popup mode may not work if you run this notebook on a remote computer (for example, Binder).
If you do not have a webcam, you can still run this demo with a video
file. Any format supported by
OpenCV
will work. You can skip first N frames to fast forward video.
Run the pose estimation:
USE_WEBCAM = False
cam_id = 0
video_file = "https://storage.openvinotoolkit.org/data/test_data/videos/store-aisle-detection.mp4"
source = cam_id if USE_WEBCAM else video_file
additional_options = {"skip_first_frames": 500} if not USE_WEBCAM else {}
run_pose_estimation(source=source, flip=isinstance(source, int), use_popup=False, **additional_options)
Source ended