1. Computer Vision Tools

1.1. Keypoint Detection - MMPose Installation

Prerequisites

• Ensure you have Python 3.7+ and CUDA 9.2+ installed. To check, run:

  nvcc --version
  python --version
  

Step 1: Install Miniconda

• Download and install Miniconda.

Step 2: Create Conda Environment with PyTorch

• Open the command prompt, create a new Conda environment, and make sure PyTorch is installed:

  conda create --name mmpose python=3.8 -y
  conda activate mmpose
  conda install pytorch torchvision -c pytorch

• Note: This approach assumes you have a GPU platform.

Step 3: Install Microsoft C++ Build Tools

• Install Microsoft C++ Build Tools from the Visual Studio Installer.
• During installation, select "Desktop Development with C++" (this includes the necessary compiler).

Step 4: Install MMEngine and MMCV

• Use OpenMMLab Installation Manager (MIM) to install MMEngine and MMCV:

  pip install -U openmim
  mim install mmengine
  mim install "mmcv>=2.0.1"
  

Step 5: Install MMDetection

• Some MMPose scripts require MMDetection for human detection. Install using:

  mim install "mmdet>=3.1.0"
  

Step 6: If Necessary, Resolve Dependency Issues

• If you encounter version incompatibility issues:
  1. Use the following command to list and inspect the versions of installed packages:

     pip list | grep mm

  2. Downgrade or upgrade dependencies as required. Example:

     pip uninstall mmcv
     mim install "mmcv<2.2.0,>=2.0.0rc4"
     mim install "mmdet>=3.0.0,<3.3.0"
     

Step 7: Install MMPose

• Finally, install MMPose as Python library with:

  mim install "mmpose>=1.1.0"
  

Step 8: Run an Inference Test

• Download a congif and checkpoint file to verify the installation:

  mim download mmpose --config td-hm_hrnet-w48_8xb32-210e_coco-256x192 --dest .

• Test the installation by applying inference on an example image in Python.
• Ensure to choose the correct Python interpreter linked to the MMPose environment created in Step 2, e.g., "3.8.20 (mmpose)":

  from mmpose.apis import inference_topdown, init_model
  from mmpose.utils import register_all_modules
  
  register_all_modules()
  
  config_file = 'td-hm_hrnet-w48_8xb32-210e_coco-256x192.py'
  checkpoint_file = 'td-hm_hrnet-w48_8xb32-210e_coco-256x192-06e76c616_20220913.pth'
  model = init_model(config_file, checkpoint_file, device='cuda:0')  # or device='cpu'
  
  # Save an image with a person
  # Run the inference and print keypoints
  results = inference_topdown(model, 'demo.jpg')
  print(results)

• The code will estimate body keypoints and display the data.
• If this works, it is possible to continue with the keypoint extraction script (i.e., estimate keypoints and save data as JSON files).

1.2. Person Detection and Tracking - Supervision Installation

Prerequisites

• Ensure you have Python 3.8+ installed.

Step 1: Create Conda Environment

• Create a new Conda environment for the tracking tasks.

  conda create -n tracking
  conda activate tracking
  

Step 2: Install Necessary Libraries

• Install the libraries of interest using Conda for improved dependency management.
• PyTorch needs to be set up in order to import the YOLO model from Ultralytics. Specify both PyTorch and Nvidia channels to ensure CUDA compatibility.

  conda install -c conda-forge supervision
  conda install conda-forge::ultralytics
  conda install pytorch torchvision torchaudio cudatoolkit=11.8 -c pytorch -c nvidia

• Activate the environment / choose correct Python interpreter when executing the tracking script.