human_motion_forecast_srnn_ros

En|中文

Summary

This repository is to realize a human motion predicion experient from paper called "Structural-RNN: Deep Learning on Spatio-Temporal Graphs" in CVPR-2016. S-RNN architecture follows the Spatio-Temporal-graph(st-graph). According to the st-graph, the spine interacts with all the body parts, and the arms and legs interact with each other. The st-graph is automatically transformed to S-RNN.

See their project page for more infomation: Structural-RNN: Deep Learning on Spatio-Temporal Graphs

Quickstart

Requirements

It is recommended to install python requirements in a virtual environment created by conda.

• ROS (Kinetic Kame on Ubuntu 16.04 or Melodic Morenia on Ubuntu 18.04)

  See the official install guide to learn how to install ROS.

• Python (2.7)

• Theano (>=0.6)

• matplotlib

• Neural Models (https://github.com/asheshjain399/NeuralModels)

Download dataset and pre-trained models

• H3.6m
• pre-trained models

You may need to create folders to to put this files.

Build and run the demo

• Clone the project code

  > git clone https://github.com/kafe6/human_motion_forecast_srnn_ros.git src/human_motion_forecast_srnn

• Build

  > catkin_make catkin_make -DCATKIN_WHITELIST_PACKAGES="human_motion_forecast_srnn" -D-DCMAKE_BUILD_TYPE=Debug

• Run the forecast node, wait it completely loaded because it may take a long time. "cp_path" is a ros parameter of your checkpoint file path.

  > roslaunch human_motion_forecast_srnn forecast.launch cp_path:=YOUR_CHECKPOINT_PATH/checkpoint.pik

• Run the publisher after the forecast node has loaded the checkpoint. "ds_path" is a ros parameter of your motion dataset path.

  > roslaunch human_motion_forecast_srnn publisher.launch ds_path:=YOUR_DATASET_PATH/walking_1.txt

• Run the visualization node to see the predicted result in rviz.

  > roslaunch human_motion_forecast_srnn fc_visualize.launch

Then you can see rviz run, and one skeletos walking in it.