Keras implementation of CNN, DeepConvLSTM, and SDAE and LightGBM for sensor-based Human Activity Recognition (HAR).

This repository contains keras (tensorflow.keras) implementation of Convolutional Neural Network (CNN) [1], Deep Convolutional LSTM (DeepConvLSTM) [1], Stacked Denoising AutoEncoder (SDAE) [2], and Light GBM for human activity recognition (HAR) using smartphones sensor dataset, UCI smartphone [3].

Table 1. The summary of the results amongst five methods on UCI smartphone dataset.

Dockerfile creates a virtual environment with Keras (tensorflow 2.3) and Python 3.8. If you wan to use this implementation locally within a docker container:

$ git clone git@github.com:takumiw/Deep-Learning-for-Human-Activity-Recognition.git
$ cd Deep-Learning-for-Human-Activity-Recognition
$ make start-gpu
# poetry install

Then run code by:

# poetry run python run_cnn.py

The dataset includes types of movement amongst three static postures (STANDING, SITTING, LYING) and three dynamic activities (WALKING, WALKING DOWNSTAIRS, WALKING UPSTAIRS).

Watch video
Download dataset

To represent raw sensor signals as a feature vector, 621 features are created by preprocessing, fast Fourier transform (FFT), statistics, etc. (run_generate_features.py.)
The 621 features of the training dataset were trained on a LightGBM classifier by 5-fold cross-validation, and evaluated on the test dataset.

The preprocessed raw sensor signals are classified CNN. The CNN architecture, which is a baseline to DeepConvLSTM, follows the study [1]

The preprocessed raw sensor signals are classified DeepConvLSTM. The fully connected layers of CNN are replaced with LSTM layers. The DeepConvLSTM architecture follows the study [1].

The preprocessed raw sensor signals are trained with SDAE, softmax layer is superimposed on top of Encoder, then whole network is fine-tuned for the target classification task. I used as same settings to the reference [2] as possible, but could not reproduce the result in the paper.

The preprocessed raw sensor signals are trained with MLP, which consists of two hidden layers.

[1] Ordóñez, F. J., & Roggen, D. (2016). Deep convolutional and lstm recurrent neural networks for multimodal wearable activity recognition. Sensors, 16(1), 115.
[2] Gao, X., Luo, H., Wang, Q., Zhao, F., Ye, L., & Zhang, Y. (2019). A human activity recognition algorithm based on stacking denoising autoencoder and lightGBM. Sensors, 19(4), 947.
[3] Reyes-Ortiz, J. L., Oneto, L., Samà, A., Parra, X., & Anguita, D. (2016). Transition-aware human activity recognition using smartphones. Neurocomputing, 171, 754-767.

• Takumi Watanabe (takumiw)
• Sophia University