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.
-
Set global configuration config.py You can modify YOLO-POSE model
POSE_MODEL
, filtersUSE_KALMAN
MINIMAL_FILTER
OUTLIER_FILTER
, output filenameTASK_NAME
, save modeSAVE_YOLO_IMG
. -
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) -
The results will be saved in
data/piclke
as.pkl
withTASK_NAME
. -
If you want to generate a demo, run
plot/plot_pickle.py
and you will get Matplotlib figures generated indata/figure
. -
Then run
plot/creat_video_from_img.py
and get video demo indata/video
. Before generate the video, you can select desired frame interval inconfig
. NOTE: The opencv and matplotlib might conflict because of PyQt5 and cause dump conflicts, try to avoidimport
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.