coloriz / HAR-WISDM_ar

Human Activity Recognition - WISDM's activity prediction dataset

Human Activity Recognition - WISDM's activity prediction dataset

Model A (1D ConvNet)

Model B (1D ConvNet + LSTM)

With a widespread of sensor-rich mobile devices, the analysis of human activities becomes more general and simpler than ever before. In this paper, we propose two deep neural networks that efficiently and accurately perform human activity recognition (HAR) using tri-axial accelerometers. In combination with powerful modern deep learning techniques like batch normalization and LSTM networks, our model outperforms baseline approaches and establishes state-of-the-art results on WISDM dataset.

This research is supported by Ministry of Culture, Sports and Tourism(MCST) and Korea Creative Content Agency(KOCCA) in the Culture Technology(CT) Research & Development Program 2019.

Quick Start

Prerequisites

• Python 3.7
• tensorflow >= v2.2
• tensorflow-datasets
• matplotlib
• mlxtend
• numpy
• pandas
• scikit-learn

Data Pre-processing

$ ./download_dataset.sh
$ python preprocess.py

Training

To train Model A,

$ python train.py ModelA

To train Model B,

$ python train.py ModelB

Each model will be saved in model/ModelA and model/ModelB respectively.

Evaluation

Confusion matrix is an useful way of representation to visualize the accuracy of which our model predicted on individual classes. To compute and plot a confusion matrix,

$ python test_models.py

or you can plot training/eval losses and accuracy over epoch using TensorBoard.

$ tensorboard --logdir=log

Appendix

Sample data (Jogging)

Confusion matrix

• (a) Model A
• (b) Model B