ADL_FEN_FLN

SCAI Semester Project - Robust Classification of the Activities of Daily Living using Unobtrusive Sensors and Transfer Learning

Overview

The goal is to effectively classify human activities. The study introduces a novel pipeline capable of transforming any kinematic signal associated with human motion into a robust feature representation. We outperformed numerous state-of-the-art models across 11 public Human Activity Recognition (HAR) datasets. Through comprehensive comparisons with baseline models on multiple datasets, we established the superiority of our proposed pipeline. Notably, our investigation extends beyond existing benchmarks by evaluating our pipeline on sensei-v2 dataset. This dataset not only represents a distinct population but also incorporates diverse kinematic modalities. The results underscore the pipeline’s remarkable generalization capabilities in feature representation of human movement.

Run Program

Initialize the Working Environment

To be able to run this program, the user must first set up a conda environment. The user must have Anaconda installed on their device. To create a new environment, run the following command on a terminal:

conda env create --name bp_estimation

Activate the created environment by running the following command:

conda activate ADL_FEN_FLN

The user must install the correct packages to run the Jupyter notebooks with the command:

conda install <PACKAGE>

The user can also directly create an environment by running the command:

conda env create -f requirements.yml

The user will have to start a Jupyter notebook server to be able to run the .ipynb files. This is done either by openning the application of choice (the application used for coding was Visual Studio Code) or the Jupyter Notebook browser of Anaconda with:

jupyter notebook

Data Preprocessing

The 'Data_Generated.ipynb' does the data preprocessing for all 11 public HAR datasets. Each signal will be one column and the last two columns are subjects and labels. All the datasets will be resampled to 20Hz with standardization. Each dataset will be saved to a pickle file.

Best Hyper-parameters

The 'Dataset_params.py' shows all the hyper-parameters for 11 public datasets. It is from hyper-parameter tuning and be chosen with best classification results for each dataset.

Model

The 'Model.py' is the Feature Extraction Network (FEN) and Feature Learning Network (FLN) model.

Training Process

The 'Train.py' is the training process for all 11 public datasets.

Transfer learning

The 'Transfer_learning.py' uses transfer learning on the sensei-v2 dataset with Best saved FEN weights to get classification results.

Directory Layout

Experiment
|
|___Codes 
|       |
|       |___Data_Generated.ipynb
|       |___Dataset_params.py
|       |___Model.py
|       |___Train.py
|       |___Transfer_learning.py
|               
|___ReadMe.md 

Description of files

Non Python files:

Filename	Description
ReadMe.md	Text file (markdown format), description of the project

Python files:

Filename	Description
Data_Generated.ipynb	For data pre-processing
Dataset_params.py	For hyper-parameters
Model.py	For FEN and FLN model
Train.py	For training process
Transfer_learning.py	For transfer learning on sensei-v2 dataset