Hidden Markov Model in Human Activity (Standing, Walking, Running) Classification

Table of Contents

• Introduction
• Installation
• Usage and Structure
• Key Algorithms
• Message to Peer Reviewers

Introduction

This repository hosts the code for the project of Pattern Recognition and Machine Learning course. We use HMM to classify the accelerometer data sequence (3 axis) recorded from three different human activities (Standing, Walking, Running). The best approach achieves 97.41% classification accuracy in validation data.

Installation

To set up the project environment, clone this repository and install the required dependencies:

git clone https://github.com/cloudisyzy/EQ2341-Hidden-Markov-Model.git
pip install -r requirements.txt

Usage and Structure

• train.ipynb: Train HMM using training data and EM, some key features of trained HMM are displayed.
• test_classification.ipynb: Use the trained HMM to classify short samples from the three classes. The short samples are not involved in training. Plot the classification accuracy and confusion matrix
• test_state_prediction.ipynb: Use the trained HMM to predict the state sequence of two long samples, which are also not involved in training. Plot the predicted state sequences.
• visualize_data.ipynb: Display the accelerometer data in 2D and 3D fashion. Readers can ignore this script if error arises, possibly due to javascript configurations.

For all .ipynb files, you don't need to change any variables, just RUN each block one by one, inside the scripts there are also short comments that helps to understand.

• data/: Data collected by the authors are stored in this folder. train.csv is used to train HMM. test_1.csv and test_2.csv are used in validating the accuracy of predicted state of long sequences. data/running_test, data/walking_test, data/standing_test consists of short samples of various lengths of accelerometer data from three activities, which are used in validating the classification accuracy of HMM.
• images/: Stores the images used in presentation.
• PattRecClasses/: Consists of core classes and functions in this project, including the basic HMM structure implemented in Assignment 1 and 2. func.py stores most of the newly defined functions in this project, including the BaumWelch Algorithm to train HMM.
• weights/: Store the pretrained weights of HMMs, the author prefer hmm_0511_wnd5_epoch30.pkl, which means training such a HMM use 30 iterations of BaumWelch algorithm, and a 5-order moving average is applied for data pre-processing.

EM (or BaumWelch) is a unsupervised method, one cannot accurately control which hidden state corresponds to which pattern. In this project case, I use the single train.csv file to train a three-state HMM.

• It turns out that "State 1" corresponds to "Standing"; "State 2" corresponds to "Walking"; "State 3" corresponds to "Running"

Key Algorithms

Expectation Maximization, or Baum-Welch Algorithm. Corresponds to PattRecClasses.func.BaumWelch, used to train HMM. Making Prediction using trained HMM, via Viterbi Algorithm or Gamma. Corresponds to PattRecClasses.func.HMM_prediction, used to classifying sequence or output the state sequence of given observations.

Message to Peer Reviewers

This README file is written in Markdown format. For the rendered version, you can either open it with IDEs, or copy and paste the whole content to https://dillinger.io/
For reviewing my project, I think run all the notebooks, .ipynb files, under the root directory is enough, which involves both training and validation of my HMM performance. There are short comments that help to understand in each file. You can also refer to Presentation.pptx for the methodology and analysis of this project.
This is a small project so I primarily use Jupyter Notebooks (.ipynb) to conduct simulations. If you are unfarmiliar with this, you can also convert the notebooks to .py files using

jupyter nbconvert --to script *.ipynb

But I cannot guarantee all scripts can be executed with no error in .py format. thx :-)