Official implementation of "STCAE"

This is the implementation of our paper "Spatio-Temporal Convolutional Autoencoders for Perimeter Intrusion Detection" (RRPR 2021). We propose 3D convolutional autoencoder for Fall detectoion and Intrusion detection task.

Installation

1. Install conda from https://docs.conda.io/en/latest/miniconda.html depending on your OS.

2. Create conda environment from the given environment.yml file.

   Go to root location of this project in your terminal and run the following command

   conda env create -f environment.yml

3. If there are errors, proceed with env.txt file.

   Run the following command

   conda create --name stcae --file env.txt

4. Activate conda environment

   source activate stcae

Code Usage:

The code base is split into two main subsets

stcae_train.py - For training different models

stcae_test.py - For testing different models

Training:

To use this code, first run the training module as:

python3 stcae_train.py

A model is then saved to Models/Dataset/....

Specify in stcae_train.py:

Dataset (Task) - dset = 'Thermal_Intrusion' or 'Thermal_Fall' or 'Thermal_Dummy'

Model - Upsampling, Deconvolution or C3D

Number of epochs - 500 by default

and other parameters.

Testing:

To evaluate the model, run the test module as:

python3 stcae_test.py

The results of testing will be saved to AEComparisons. Once training has completed, find the saved model under Models/Thermal/{model_name}. To evaluate the model, set the variable pre_load to the path to this model. Run stcae_test.py and find the results in AEComparisons. The Labels.csv file under each dataset provides the ground truth labels for start and end of fall frames.

Generating Animation:

To generate an animation:

run stcae_test.py, with animate option set to True.

An animation (mp4 file) for each testing video will be saved to Animation folder.

Requirements:

Keras - 2.2.2
Tensorflow - 1.10.0
Python - 3.6.4

Dataset Sharing:

Please contact authors of DeepFall paper for access to preprocessed data for Fall Detection.

For intrusion detection, we have a private dataset which unfortunately we cannot share.

Place the data in folder Datasets. See README.txt in Datasets for information on using the shared data.

Results: All results are in AEComparisons folder.

Illustrations:

1) Fall Detection

2) Intrusion Detection

Bibtex

@inproceedings{lohani2021spatio,
  title={Spatio-temporal convolutional autoencoders for perimeter intrusion detection},
  author={Lohani, Devashish and Crispim-Junior, Carlos and Barth{\'e}lemy, Quentin and Bertrand, Sarah and Robinault, Lionel and Tougne, Laure},
  booktitle={International Workshop on Reproducible Research in Pattern Recognition},
  pages={47--65},
  year={2021},
  organization={Springer}
}