Human Pose Detection on EdgeTPU

Overview

Coral PoseNet

Pose estimation refers to computer vision techniques that detect human figures in images and video, so that one could determine, for example, where someone’s elbow, shoulder or foot show up in an image. PoseNet does not recognize who is in an image, it is simply estimating where key body joints are.

This repo contains a set of PoseNet models that are quantized and optimized for use on Coral's Edge TPU, together with some example code to shows how to run it on a camera stream.

Why PoseNet ?

Pose estimation has many uses, from interactive installations that react to the body to augmented reality, animation, fitness uses, and more. We hope the accessibility of this model inspires more developers and makers to experiment and apply pose detection to their own unique projects, to demonstrate how machine learning can be deployed in ways that are anonymous and private.

How does it work ?

At a high level pose estimation happens in two phases:

  1. An input RGB image is fed through a convolutional neural network. In our case this is a MobileNet V1 architecture. Instead of a classification head however, there is a specialized head which produces a set of heatmaps (one for each kind of key point) and some offset maps. This step runs on the EdgeTPU. The results are then fed into step 2)

  2. A special multi-pose decoding algorithm is used to decode poses, pose confidence scores, keypoint positions, and keypoint confidence scores. Note that unlike in the TensorflowJS version we have created a custom OP in Tensorflow Lite and appended it to the network graph itself. This CustomOP does the decoding (on the CPU) as a post processing step. The advantage is that we don't have to deal with the heatmaps directly and when we then call this network through the Coral Python API we simply get a series of keypoints from the network.

If you're interested in the gory details of the decoding algorithm and how PoseNet works under the hood, I recommend you take a look at the original research paper or this medium post whihch describes the raw heatmaps produced by the convolutional model.

Important concepts

Pose: at the highest level, PoseNet will return a pose object that contains a list of keypoints and an instance-level confidence score for each detected person.

Keypoint: a part of a person’s pose that is estimated, such as the nose, right ear, left knee, right foot, etc. It contains both a position and a keypoint confidence score. PoseNet currently detects 17 keypoints illustrated in the following diagram:

pose keypoints

Keypoint Confidence Score: this determines the confidence that an estimated keypoint position is accurate. It ranges between 0.0 and 1.0. It can be used to hide keypoints that are not deemed strong enough.

Keypoint Position: 2D x and y coordinates in the original input image where a keypoint has been detected.

Examples in this repo

NOTE: PoseNet relies on the latest Pycoral API, tflite_runtime API, and libedgetpu1-std or libedgetpu1-max:

Please also update your system before running these examples. For more information on updating see:

To install all other requirements for third party libraries, simply run

sh install_requirements.sh

simple_pose.py

A minimal example that simply downloads an image, and prints the pose keypoints.

python3 simple_pose.py

pose_camera.py

A camera example that streams the camera image through posenet and draws the pose on top as an overlay. This is a great first example to run to familiarize yourself with the network and its outputs.

Run a simple demo like this:

python3 pose_camera.py

If the camera and monitor are both facing you, consider adding the --mirror flag:

python3 pose_camera.py --mirror

In this repo we have included 3 posenet model files for differnet input resolutions. The larger resolutions are slower of course, but allow a wider field of view, or further-away poses to be processed correctly.

posenet_mobilenet_v1_075_721_1281_quant_decoder_edgetpu.tflite
posenet_mobilenet_v1_075_481_641_quant_decoder_edgetpu.tflite
posenet_mobilenet_v1_075_353_481_quant_decoder_edgetpu.tflite

You can change the camera resolution by using the --res parameter:

python3 pose_camera.py --res 480x360  # fast but low res
python3 pose_camera.py --res 640x480  # default
python3 pose_camera.py --res 1280x720 # slower but high res

anonymizer.py

A fun little app that demonstrates how Coral and PoseNet can be used to analyze human behavior in an anonymous and privacy-preserving way.

Posenet converts an image of a human into a mere skeleton which captures its position and movement over time, but discards any precisely identifying features and the original camera image. Because Coral devices run all the image analysis locally, the actual image is never streamed anywhere and is immediately discarded. The poses can be safely stored or analysed.

For example a store owner may want to study the bahavior of customers as they move through the store, in order to optimize flow and improve product placement. A museum may want to track which areas are most busy, at which times such as to give guidance which exhibits may currently have the shortest waiting times.

With Coral this is possible without recording anybody's image directly or streaming data to a cloud service - instead the images are immediately discarded.

The anaonymizer is a small app that demonstrates this is a fun way. To use the anonymizer set up your camera in a sturdy position. Lauch the app and walk out of the image. This demo waits until no one is in the frame, then stores the 'background' image. Now, step back in. You'll see your current pose overlayed over a static image of the background.

python3 anonymizer.py

(If the camera and monitor are both facing you, consider adding the --mirror flag.)

video of three people interacting with the anonymizer demo

synthesizer.py

This demo allows people to control musical synthesizers with their arms. Up to 3 people are each assigned a different instrument and octave, and control the pitch with their right wrists and the volume with their left wrists.

You'll need to install FluidSynth and a General Midi SoundFont:

apt install fluidsynth fluid-soundfont-gm
pip3 install pyfluidsynth

Now you can run the demo like this:

python3 synthesizer.py

The PoseEngine class

The PoseEngine class (defined in pose_engine.py) allows easy access to the PoseNet network from Python, using the EdgeTPU API.

You simply initialize the class with the location of the model .tflite file and then call DetectPosesInImage, passing a numpy object that contains the image. The numpy object should be in int8, [Y,X,RGB] format.

A minimal example might be:

from tflite_runtime.interpreter import Interpreter
import os
import numpy as np
from PIL import Image
from PIL import ImageDraw
from pose_engine import PoseEngine


os.system('wget https://upload.wikimedia.org/wikipedia/commons/thumb/3/38/'
          'Hindu_marriage_ceremony_offering.jpg/'
          '640px-Hindu_marriage_ceremony_offering.jpg -O /tmp/couple.jpg')
pil_image = Image.open('/tmp/couple.jpg').convert('RGB')
engine = PoseEngine(
    'models/mobilenet/posenet_mobilenet_v1_075_481_641_quant_decoder_edgetpu.tflite')
poses, _ = engine.DetectPosesInImage(pil_image)

for pose in poses:
    if pose.score < 0.4: continue
    print('\nPose Score: ', pose.score)
    for label, keypoint in pose.keypoints.items():
        print('  %-20s x=%-4d y=%-4d score=%.1f' %
              (label, keypoint.point[0], keypoint.point[1], keypoint.score))

To try this, run

python3 simple_pose.py

And you should see an output like this:

Inference time: 14 ms

Pose Score:  0.60698134
  NOSE                 x=211  y=152  score=1.0
  LEFT_EYE             x=224  y=138  score=1.0
  RIGHT_EYE            x=199  y=136  score=1.0
  LEFT_EAR             x=245  y=135  score=1.0
  RIGHT_EAR            x=183  y=129  score=0.8
  LEFT_SHOULDER        x=269  y=169  score=0.7
  RIGHT_SHOULDER       x=160  y=173  score=1.0
  LEFT_ELBOW           x=281  y=255  score=0.6
  RIGHT_ELBOW          x=153  y=253  score=1.0
  LEFT_WRIST           x=237  y=333  score=0.6
  RIGHT_WRIST          x=163  y=305  score=0.5
  LEFT_HIP             x=256  y=318  score=0.2
  RIGHT_HIP            x=171  y=311  score=0.2
  LEFT_KNEE            x=221  y=342  score=0.3
  RIGHT_KNEE           x=209  y=340  score=0.3
  LEFT_ANKLE           x=188  y=408  score=0.2
  RIGHT_ANKLE          x=189  y=410  score=0.2

Owner
google-coral
Open source projects for coral.ai
google-coral
Gradient representations in ReLU networks as similarity functions

Gradient representations in ReLU networks as similarity functions by Dániel Rácz and Bálint Daróczy. This repo contains the python code related to our

1 Oct 08, 2021
Toolchain to build Yoshi's Island from source code

Project-Y Toolchain to build Yoshi's Island (J) V1.0 from source code, by MrL314 Last updated: September 17, 2021 Setup To begin, download this toolch

MrL314 19 Apr 18, 2022
MicRank is a Learning to Rank neural channel selection framework where a DNN is trained to rank microphone channels.

MicRank: Learning to Rank Microphones for Distant Speech Recognition Application Scenario Many applications nowadays envision the presence of multiple

Samuele Cornell 20 Nov 10, 2022
The fastai deep learning library

Welcome to fastai fastai simplifies training fast and accurate neural nets using modern best practices Important: This documentation covers fastai v2,

fast.ai 23.2k Jan 07, 2023
Implements the training, testing and editing tools for "Pluralistic Image Completion"

Pluralistic Image Completion ArXiv | Project Page | Online Demo | Video(demo) This repository implements the training, testing and editing tools for "

Chuanxia Zheng 615 Dec 08, 2022
A PyTorch Reimplementation of TecoGAN: Temporally Coherent GAN for Video Super-Resolution

TecoGAN-PyTorch Introduction This is a PyTorch reimplementation of TecoGAN: Temporally Coherent GAN for Video Super-Resolution (VSR). Please refer to

165 Dec 17, 2022
TLDR; Train custom adaptive filter optimizers without hand tuning or extra labels.

AutoDSP TLDR; Train custom adaptive filter optimizers without hand tuning or extra labels. About Adaptive filtering algorithms are commonplace in sign

Jonah Casebeer 48 Sep 19, 2022
Nest - A flexible tool for building and sharing deep learning modules

Nest - A flexible tool for building and sharing deep learning modules Nest is a flexible deep learning module manager, which aims at encouraging code

ZhouYanzhao 41 Oct 10, 2022
Face Mask Detection system based on computer vision and deep learning using OpenCV and Tensorflow/Keras

Face Mask Detection Face Mask Detection System built with OpenCV, Keras/TensorFlow using Deep Learning and Computer Vision concepts in order to detect

Chandrika Deb 1.4k Jan 03, 2023
In the case of your data having only 1 channel while want to use timm models

timm_custom Description In the case of your data having only 1 channel while want to use timm models (with or without pretrained weights), run the fol

2 Nov 26, 2021
A Blender python script for getting asset browser custom preview images for objects and collections.

asset_snapshot A Blender python script for getting asset browser custom preview images for objects and collections. Installation: Click the code butto

Johnny Matthews 44 Nov 29, 2022
A library for augmentation of a YOLO-formated dataset

YOLO Dataset Augmentation lib Инструкция по использованию этой библиотеки Запуск всех файлов осуществлять из консоли. GoogleCrawl_to_Dataset.py Это ск

Egor Orel 1 Dec 10, 2022
Extracting and filtering paraphrases by bridging natural language inference and paraphrasing

nli2paraphrases Source code repository accompanying the preprint Extracting and filtering paraphrases by bridging natural language inference and parap

Matej Klemen 1 Mar 09, 2022
The repository for the paper "When Do You Need Billions of Words of Pretraining Data?"

pretraining-learning-curves This is the repository for the paper When Do You Need Billions of Words of Pretraining Data? Edge Probing We use jiant1 fo

ML² AT CILVR 19 Nov 25, 2022
PyTorch Implementation of Exploring Explicit Domain Supervision for Latent Space Disentanglement in Unpaired Image-to-Image Translation.

DosGAN-PyTorch PyTorch Implementation of Exploring Explicit Domain Supervision for Latent Space Disentanglement in Unpaired Image-to-Image Translation

40 Nov 30, 2022
Hack Camera, Microphone, Location, Clipboard With Just a Link. Also, Get Many Details About Victim's Device. And So On...

An Automated Tool to Hack Victim's Camera, Microphone, Location, Clipboard. Has 2 Extra Features. Version 1.1 Update Fixed Some Major Bugs Data Saving

ToxicNoob 36 Jan 07, 2023
Tooling for the Common Objects In 3D dataset.

CO3D: Common Objects In 3D This repository contains a set of tools for working with the Common Objects in 3D (CO3D) dataset. Download the dataset The

Facebook Research 724 Jan 06, 2023
Implementation of ViViT: A Video Vision Transformer

ViViT: A Video Vision Transformer Unofficial implementation of ViViT: A Video Vision Transformer. Notes: This is in WIP. Model 2 is implemented, Model

Rishikesh (ऋषिकेश) 297 Jan 06, 2023
BBScan py3 - BBScan py3 With Python

BBScan_py3 This repository is forked from lijiejie/BBScan 1.5. I migrated the fo

baiyunfei 12 Dec 30, 2022
ViViT: Curvature access through the generalized Gauss-Newton's low-rank structure

ViViT is a collection of numerical tricks to efficiently access curvature from the generalized Gauss-Newton (GGN) matrix based on its low-rank structure. Provided functionality includes computing

Felix Dangel 12 Dec 08, 2022