Deploy your apps to a supercloud in a few clicks

This Engineering Education program is supported by Section. Instantly deploy your GitHub apps, Docker containers or K8s namespaces to a supercloud.

Try It For Free

Multi-Person Pose Estimation with Python

October 26, 2021

The Multi-Person Pose Estimation is a model designed for sports and exercise-based applications. The model can detect six people in an image or video simultaneously, and 17 key-point joints per person.

This tutorial will use a technique known as Multi-Person Pose Estimation to predict human joint locations of people in an image frame.

One of the most popular use cases is in kinematics, which analyzes the body and all the different angles and joints within a body.

Table of contents


To understand this tutorial, you’ll need to be familiar with:

  • Machine learning.
  • OpenCV library.
  • Python programming language.
  • Jupyter Notebook or Google Colab. To follow through, please use Google Colab.
  • TensorFlow and TensorFlow Hub. TensorFlow Hub is a repository consisting of trained machine learning models which you can download and use in any of your projects.

You’ll need to download the footballers.mp4 video file from this Google drive account and upload it into your Google Colab. I uploaded it on my Google Colab for this tutorial but it get’s deleted after runtime. So, you’ll need to download it and upload it yourself to avoid errors.


Multi-Person Pose Estimation is one of Google’s latest models built on top of their lightning model. It is an improvement from the MoveNet.SinglePose model, which could only detect single poses. The model comprises three models.

The backbone network is built using the MobileNetV2 network, the feature extraction layer is comprised of Feature Pyramid Networks, and the layer that outputs the key points is comprised of the CenterNet model.

The Multi-Person Pose Estimation is a model designed for sports and exercise-based applications. It is intended to be used between three to six feet from your camera or webcam. If you are working with distances more than six feet, this model might not be the most appropriate for you.

The model can only detect six people in an image/video simultaneously, and 17 keypoint joints per person, i.e., nose, left eye, right eye, right ear, left ear, etc.

Some of the use cases for this model are:

  • To determine the correctness of poses of yoga class members during a yoga class.
  • To establish which swimmers have the best swimming strokes and those who need improvements during swimming lessons.
  • During supervision of a tennis lesson. The model would help analyze the people with correct and incorrect stroke positions.
  • In medical contexts to identify abnormal poses in patients.

To use the Multi-Person Pose Estimation model, we will need to perform three processes:

  1. Install TensorFlow and TensorFlow Hub.
  2. Load the Multi-Person Movenet lightning model.
  3. Make detection from videos.

Let’s get started!

Installing and importing dependencies

!pip install tensorflow==2.4.1 tensorflow-gpu==2.4.1 tensorflow-hub opencv-python matplotlib
import tensorflow as tf
import tensorflow_hub as hub
import cv2
from matplotlib import pyplot as plt
import numpy as np
  • numpy will be used to help draw the keypoints and edges
  • cv2 will allow us to leverage the OpenCV library for computer vision. In this tutorial, it will help us when using the webcam and running videos.
  • tensorflow_hub will allow us to use the pre-trained Multi-Person Movenet model.

If you’re using a GPU, consider limiting the memory growth of your GPU to avoid receiving out-of-memory errors. This restricts TensorFlow from using all the V-RAM on your machine.

gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
    tf.config.experimental.set_memory_growth(gpu, True)

Loading the model from TF Hub

model = hub.load('')
movenet = model.signatures['serving_default']

If you’ve read the model’s documentation, you’ll realize that these two lines of code have been fetched directly from it.

The first variable, model is downloading the model, while the second variable, movement is setting up a new variable to go ahead and use that model.

Making detection using video or webcam

cap = cv2.VideoCapture(0)
while cap.isOpened():
        ret, frame =

        cv2.imshow('Multi-Person Pose Estimation', frame)

        if cv2.waitKey(10) & 0xFF==ord('q'):

The code above is a standard OpenCV code that we’ve used in previous OpenCV tutorials. If you’re not familiar with it, you can read this documentation.

The code establishes a connection to our webcam, reads the frames in our images/videos, and renders the frames on our screen. The last piece of code shows us how to exit our application when done.

Note that the value 0 in VideoCapture(0) may change depending on the number of video devices you have. If 0 is not working, you can try and change this number.

We will pass our frame captured to the model to use the frames with our model.

This is done as follows:

from google.colab.patches import cv2_imshow

cap = cv2.VideoCapture('footballers.mp4')
while cap.isOpened():
    ret, frame =
    # Resize image
    img = frame.copy()
    img = tf.image.resize_with_pad(tf.expand_dims(img, axis=0), 384,640)
    input_img = tf.cast(img, dtype=tf.int32)
    # Detection section
    results = movenet(input_img)
    keypoints_with_scores = results['output_0'].numpy()[:,:,:51].reshape((6,17,3))
    # Render keypoints 
    loop_through_people(frame, keypoints_with_scores, EDGES, 0.1) #0.1 is our confidence_threshold
    if cv2.waitKey(10) & 0xFF==ord('q'):

On the documentation, to run the model inference on an input image, it is recommended to resize and pad your image (adding 0s if the size is odd) so that:

  • The height/width are both multiples of 32.
  • The height to width ratio should be close enough to cover the original image’s aspect ratio.
  • The larger side should be made to be 256px. For example, a 720p image (i.e. 720x1280 (HxW)) should be resized and padded to 160x256 image.

This is what that additional code does. We then make our detections using the movenet model and finally render the keypoints.

The output is a float32 tensor of shape [1, 6, 56]. 1 represents the batch dimension, 6 represents the maximum number of people the model can detect, and 56 represents the predicted bounding box/keypoint locations and scores.

The first 51 elements (17*3) are the keypoint locations and scores. The last five elements represent the region of the bounding box and the confidence score of the instance.

We’ve used NumPy [:,:,:51] to grab all batch dimensions, all of the maximum number of people, and only the 51 keypoint locations and scores because that’s all we need to render our image. reshape((6,17,3) tells us that we’re reshaping the keypoint to have 6 people, each person with 17 joints, each person having 3 values; the x,y, and confidence score.

The order of the 17 keypoint joints is: [nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, right ankle].

Below is the function to loop through each person detected and renders.

def loop_through_people(frame, keypoints_with_scores, edges, confidence_threshold):
    for person in keypoints_with_scores:
        draw_connections(frame, person, edges, confidence_threshold)
        draw_keypoints(frame, person, confidence_threshold)

The confidence_threshold tells the model to draw only keypoints and edges if the threshold is reached. For example, if the confidence score is 0.2 but the confidence_score is set to 0.5, that joint will not be drawn.

All that’s remaining is to draw the keypoints and edges.

Drawing keypoints and edges

We’ll create the draw_keypoints and draw_connections which will help us render our videos after detection.

def draw_keypoints(frame, keypoints, confidence_threshold):
    y, x, c = frame.shape
    shaped = np.squeeze(np.multiply(keypoints, [y,x,1]))
    for kp in shaped:
        ky, kx, kp_conf = kp
        if kp_conf > confidence_threshold:
  , (int(kx), int(ky)), 6, (0,255,0), -1)
def draw_connections(frame, keypoints, edges, confidence_threshold):
    y, x, c = frame.shape
    shaped = np.squeeze(np.multiply(keypoints, [y,x,1]))
    for edge, color in edges.items():
        p1, p2 = edge
        y1, x1, c1 = shaped[p1]
        y2, x2, c2 = shaped[p2]
        if (c1 > confidence_threshold) & (c2 > confidence_threshold):      
            cv2.line(frame, (int(x1), int(y1)), (int(x2), int(y2)), (0,0,255), 4)
    (0, 1): 'm',
    (0, 2): 'c',
    (1, 3): 'm',
    (2, 4): 'c',
    (0, 5): 'm',
    (0, 6): 'c',
    (5, 7): 'm',
    (7, 9): 'm',
    (6, 8): 'c',
    (8, 10): 'c',
    (5, 6): 'y',
    (5, 11): 'm',
    (6, 12): 'c',
    (11, 12): 'y',
    (11, 13): 'm',
    (13, 15): 'm',
    (12, 14): 'c',
    (14, 16): 'c'

The values above tell us how to connect the joints. Remember we mentioned the order in which the 17 joints are read. So, for example, the first value, (0, 1): 'm', tells us how the nose connects to the left eye, and the last value, (14, 16): 'c' tells us how the right knee connects to the right ankle.

Please find all the code for this tutorial here.

Wrapping up

To recap, we installed and imported our dependencies, loaded the model, made detections using either the webcam or videos and finally drew the keypoints and edges.

If you’re not getting accurately drawn keypoints and edges, or it’s drawing on objects that it shouldn’t, you can try tuning the and confidence score and scale value to obtain the correct aspect ratio.

The larger the image that you pass through, the slower the detections are going to be. The opposite is also true.

Happy coding!

Further reading

Peer Review Contributions by: Collins Ayuya