---
title: "Realtime human pose estimation with TensorFlow and PoseNet"
language: en
topics: ["art", "data", "experiment", "game", "people", "sensors", "work"]
full_article_url: https://willem.com/en/2019-12-01_realtime-human-pose-recognition-through-computer-vision/
---
# Realtime human pose estimation with TensorFlow and PoseNet

*How a pretrained model finds body joints in a video feed.*

> PoseNet is a TensorFlow vision model that estimates a person's pose by locating key body joints, like elbows, hands, hips and knees. When I explored it in 2019, it ran in realtime on ordinary hardware because the model was already trained.

TensorFlow is an open source machine learning platform originally developed by Google. A tensor is a multi-dimensional array of numeric values, and images flow through a chain of steps, cropping, resizing, removing colour, comparing contours, from unanalysed to recognised.
The heavy lifting happens when models are trained, so running a trained model does not need very expensive hardware. PoseNet is such a pretrained vision model: it estimates a person's pose by finding key body joints, elbows, hands, hips, knees and ankles.
Body detection first uses the Single Shot MultiBox Detector (SSD), a fast algorithm that shrinks the image into simplified feature maps and predicts early, which suits large objects like human bodies. With poses detected live, a person can quite literally become the game controller or drive an interactive art installation.