Efficient multi-human 3D skeleton tracking based on RealSenese RGBD

DEMO

GIF

Python 3.10

numpy, ultralytics, torch, opencv-python

Details in pyproject.toml

Poetry is recommended to initialize this repository, where pyproject.toml and poetry.lock are provided.

1. Set global configuration config.py You can modify YOLO-POSE model POSE_MODEL, filters USE_KALMAN MINIMAL_FILTER OUTLIER_FILTER, output filename TASK_NAME, save mode SAVE_YOLO_IMG.

2. Run feature_extractor/skeleton_extractor_node.py with activated ROS TOPICS. NOTE: make sure the image types in configuaration match your ROS TOPICS(Imgae vs Compressed Image)

3. The results will be saved in data/piclke as .pkl with TASK_NAME.

4. If you want to generate a demo, run plot/plot_pickle.py and you will get Matplotlib figures generated in data/figure.

5. Then run plot/creat_video_from_img.py and get video demo in data/video. Before generate the video, you can select desired frame interval in config. NOTE: The opencv and matplotlib might conflict because of PyQt5 and cause dump conflicts, try to avoid import them in the same PID.

ultralytics YOLO v8 requires all the following repository with GNU AGPLv3.

• ROS2 dev, Real-time deployment on robot and calibration with multi-view pseudo ground true.
• Descriptor and Feature
• Action prediction based motion planning.