abhinavadarsh / bidirectional_har

Using bidirectional LSTM for Human Activity Recognition

LSTM for Human Activity Recognition

LSTM based human activity recognition using smart phone sensor dataset(a cellphone attached on the waist). Classifying the type of movement amongst six categories:

• WALKING,
• WALKING_UPSTAIRS,
• WALKING_DOWNSTAIRS,
• SITTING,
• STANDING,
• LAYING.

Dataset

Dataset can be downloaded at https://archive.ics.uci.edu/ml/machine-learning-databases/00240/UCI%20HAR%20Dataset.zip

Follow this link to see a video of the how data is collected

The sensor signals (accelerometer and gyroscope) were pre-processed by applying noise filters and then sampled in fixed-width sliding windows of 2.56 sec and 50% overlap (128 readings/window). The sensor acceleration signal, which has gravitational and body motion components, was separated using a Butterworth low-pass filter into body acceleration and gravity. The gravitational force is assumed to have only low frequency components, therefore a filter with 0.3 Hz cutoff frequency was used.

Model

In this repo, we adopt a 2 layer stacked basic LSTM and use almost the raw data: only the gravity effect has been filtered out of the accelerometer as a preprocessing step for another 3D feature as an input to help learning.

Usage

1. Install TensorFlow r1.1
2. Clone this repo by git clone https://github.com/csarron/lstm_har.git
3. For training the model, use python train_lstm.py (you can pass layer, unit, training epochs as arguments)
4. For prediction, use python predict_lstm.py

Results

We got 93.48% test accuracy, it took 9186.56s for training on CPU (MacBookPro12,1; Intel Core i7, 3.1GHz; Mem 16GB DDR3 1867 MHz)

Acknowledgement

The original source code is modified from guillaume-chevalier