/Human-Motion-Prediction

Primary LanguageMATLAB

Human-Motion-Prediction

By Yujiao Cheng, Weiye Zhao, Changliu Liu

Introduction

Human-Motion-Prediction is currently implemented with two different algorithms, RLS-PAA and Identifier-based algorithm.

Requirements: software

  1. MATLAB 2014a or later.

Requirements: hardware

CPU, Windows 7 or later, MAC OS.

Fake Data

  1. Run fake_data_demo.m to generate arficial motion system training dataset.
    • Note: check opts. parameter in demo files for more parameter setting.

Demo

  1. Run id_demo.m to apply identifier-based algorithm on human motion data.
  2. Run rls_demo.m to apply RLS-PAA algorithm on human motion data.

Offline Neural Network Training

  1. .\offline_train\trainNN.py to training offline models on human motion data.
  2. human motion dataset should be set manually. Please find Kinect and CMU mocap datasets in .\data2.

Online Adaptation

  1. id_demo.m and rls_demo.m are demos for two online adaptation algorithms. Check both files for more details.
    • Note: check opts. parameter in demo files for more parameter setting.
  2. Check other scripts in ./lib for auxiliary function.

Results

  1. The results in terms of prediction error and prediction motion state of two algorithms on four datasets are stored in ./results/.., please run plot_err(error, instance number, 'y label', 'x label') to see the graph for prediction error after loading the error mat.

Note:

  • In all the experiments, online adaptation is performed on smoothed human motion data.\data2\trainX&Y. Both ready-for-adaptation trainX or trainY data are stored in .\data2\data_time.mat or .\data2\cmu_data.mat, which denote Kinect dataset and CMU mocap dataset respectively.
  • artificial system data is not stored, but users can run fake_data_demo.m to genreate user defined artificial motion data.
  • pre-stored offline trained NN initiation parameters for CMU dataset and artificial systems can be found in .\para\.., parameters for Kinect dataset can be found in .\data2.
  • Running time is not recorded, but normally id-based algorithm are slower than RLS algorithm.