kleincode/webcam-pose-estimation

Minimal real-time webcam pose estimation with OpenCV & TFLite.

Real-Time Webcam Pose Estimation using OpenCV, TFLite, and MoveNet

This is a minimal C++ project for performing real-time pose estimation on your webcam based on OpenCV, TFLite and MoveNet. It's purpose is not to be useful but to get started with OpenCV and TFLite in C++.

Features

• Webcam capture and preview using OpenCV
• Choose between MoveNet SinglePose and MoveNet Multipose
• Select input size for multi-pose model (speed vs. accuracy)
• FPS counter

Installation

Prerequisites

Requirements:

• Python 3.9+
• Conan 2.0.4+
• CMake
• PreMake
• Visual Studio 2022 (for other IDEs, adjust the premake5 command in init.py)

Installing Conan: You might want to use an Anaconda environment if your global Python installation is messed up. Make sure that you are always in the same Python environment when using Conan.

# if not yet installed
pip install conan
conan --version # should be 2.0.13 (or higher)
conan profile detect

Open the profile file in the last line of the output or run conan profile path default to see the same file path. Update the following line to 20:

compiler.cppstd=20

Install CMake if not already installed (check with cmake --version). For Windows, download the installer from https://cmake.org/download/.

Install Premake: Download Premake from https://premake.github.io/download and put the premake5.exe into the root folder or your PATH. You don't need the other files from the downloaded zip archive. Check premake5 is accessible from your console by typing premake5 --version.

Method 1: Auto-install repository (recommended)

To use the auto-install script, run:

python init.py

Make sure to use the same Python environment you installed Conan in! This usually takes a while (> 20 min), mostly because of the debug build.

You can open the generated WebcamPoseEstimation.sln file in Visual Studio 2022.

Method 2: Manual installation

1. Install packages using Conan:

Run both commands, one for the Release and one for the Debug setting:

conan install . --build missing --output-folder=./dependencies
conan install . --build missing --output-folder=./dependencies --settings=build_type=Debug

As a result, a lot of things should happen in the console and the dependencies folder should be generated and populated with .lua and .bat files. Note that the debug command usually takes a lot longer to build on first execution.

Possible reasons for failure:

• CMake is not installed or too old -> check if that is the case using cmake --version and if applicable, install CMake for your system (Windows: google and download the installer).
• You used an Anaconda Python to install Conan but are not in an Anaconda environment -> usually no response. Make sure your Anaconda environment is active.

2. Build a Visual Studio 2022 project:

premake5 vs2022

This generates the WebcamPoseEstimation.sln file, which can be opened in Visual Studio 2022. The command is safe to repeat. It should particularly be repeated after any premake5/lua files were updated.

Acknowledgements

• The Conan setup is based on Lötwig Fusel's video: https://www.youtube.com/watch?v=7sLeMVUo8Kg.
• Source code is inspired by https://github.com/conan-io/examples2/tree/main/examples/libraries/tensorflow-lite/pose-estimation