HarSkel - Human Action Recognition from Skeleton joints

This software is provided as additional material for the following paper:

Learning features combination for human action recognition from skeleton sequences

Dependencies

HarSkel is a Matlab(r) software, but it does not depend on specific toolboxes.

Horever, it needs the following libraries (which are provided in the 3rdparty folder):

• LIBSVM
• VLFeat

Before running the software, compile them (under Linux):

cd 3rdparty
./build.sh

Setting up

The file setup.m configures the environment to run the software.

Consider taking a look at this file before running it. All the parameters and the dataset to be used are configured there.

Datasets

We provide pre-computed skeleton sequences for all the datasets supported:

• MSR Action 3D
• UTKinect Action 3D
• Florence 3D Actions

If you want to regenerate them, please check in the file recomp_skeletons.m.

Running examples

In order to reproduce the results reported in the paper, run the file train_and_eval.m.

By default, the software is setted up for the MSR Action 3D dataset. The terminal output should look like this:

>> train_and_eval
pca: reduce features size from 8970 to 512
Ep. 00000 | G 14867.1 | Eta 0 | N.Imp 26423 | Loss 30829.1 | Acc 83.2%
Ep. 00001 | G 3477.25 | Eta 4.10069e-05 | N.Imp 12206 | Loss 11583.4 | Acc 89.4%
Ep. 00002 | G 311.288 | Eta 0.000237151 | N.Imp 02924 | Loss 2456.4 | Acc 92.7%
...

Citing

If this software is useful for you (or any part of it), please consider citing us:

@article{Luvizon_PRL_2017,
  author = "Diogo Carbonera Luvizon and Hedi Tabia and David Picard",
  title = "Learning features combination for human action recognition from skeleton sequences",
  journal = "Pattern Recognition Letters",
  volume = "99",
  pages = "13 - 20",
  year = "2017",
  doi = "https://doi.org/10.1016/j.patrec.2017.02.001",
}

License

MIT License