Developing a Start-to-Finish Pipeline for Accelerometer-Based Activity Recognition Using Long Short-Term Memory Recurrent Neural Networks
This repository contains the source code associated with the SciPy 2018 Proceedings paper and associated poster "Developing a Start-to-Finish Pipeline for Accelerometer-Based Activity Recognition Using Long Short-Term Memory Recurrent Neural Networks.”
Accelerometer data is read in and formatted for a Data Analysis Pipeline within the folder src/data/.
In src/models/, a baseline LSTM is optimized based on a wide range of hyperparameter settings found throughout literature (i.e., a 30-study meta-analysis style overview of the field of human activity recognition (HAR) using LSTM models).
The optimized LSTM is then incorporated in a proposed Data Analysis Pipeline intended to foster reproducibility and scientific rigor within the field. All scripts in src/models/ import the necessary functions from src/data/ to prepare the data and run the model in a single executable.
The poster and other documents are found in the folder docs/.
As of July 2018, this repo has been made to work on the UCI HAR Dataset from Reyes-Ortiz, et. al. (2013). Solely the triaxial total accelerometer signal (vs. the body-only signal with gravity component removed) as well as the UCI HMP Dataset from Bruno, et. al. (2012).
Note: If you would like help formatting another dataset for this pipeline, or if you have another issues/comments, please make an Issue and include an @xtianmcd tag.
Additionally, if you use this code for any project, please let me know at clm121@uga.edu and cite the repo, and feel free to submit a PR to upload any additional code you might write to the repo.