/Driver-drowsiness-classification

Mini-project using the PyTorch framework

Primary LanguagePython

Driver-drowsiness-classification-based-on-reaction-time

Classifiying driver drowsiness based on the dataset from the paper "Multi-channel EEG recordings during a sustained-attention driving task".

Project Topic & Goal

The goal of this project is the development of subject-dependent EEG-based driver drowsiness state (i.e., awake, drowsy) classification model.

  • Estimating driver fatigue is an important issue for traffic safety and user-centered brain-computer interface.
  • The drowsy state is caused by various factors such as fatigue, workload, and distraction.
  • Physiological features are used to assess drowsiness because they are continuously available and could be considered as an objective, more direct, measure of the mental state.

Dataset

Title: Multi-channel EEG recordings during a sustained-attention driving task

Author: Z. Cao, C.-H. Chuang, J.-K. King, and C.-T. Lin

Journal: Scientific data

Year: 2019

Original dataset: https://figshare.com/articles/dataset/Multi-channel_EEG_recordings_during_a_sustained-attention_driving_task/6427334/5

Preprocessed dataset: https://figshare.com/articles/dataset/Multi-channel_EEG_recordings_during_a_sustained-attention_driving_task_preprocessed_dataset_/7666055/3

Experimental protocol

  • 60-90 minutes simulated driving on an empty highway
  • Measure reaction time while performing a sustained-attention task
    • Steer the wheel when a participant recognizes a random lane-deviation (5-10s)
    • Reaction time (RT) = response onset - deviation onset
  • 30-channel EEG data for 27 subjects (22-28 years)
    • FP1, FP2, F7, F3, FZ, F4, F8, FT7, FC3, FCZ, FC4, FT8, T3, C3, CZ, C4, T4, TP7, CP3, CPZ, CP4, TP8, T5, P3, PZ, P4, T6, O1, OZ, O2

Data selection and labeling

  • Select 14 suhbject data
  • Data preprocessing:
    • 1-50Hz bandpass filtering for the data pre-processing
    • Down-sample to 250Hz
    • Linearly interpolate the RT between the lane-deviation happening
    • Segmentation

Method & Performance

To improve the performance of the classification, we utilized the ESTCNN with minor modifications, which was derived the higher performance in the drowsiness classification. drow_1

Experimental Setup

  1. We DO NOT change the "dataloader" for a fair comparison with baseline performance.
  2. Best model selection was performed from the balanced accuracy in the validation set.
  3. Hyperparameters:
  • Epochs: 100
  • Batch size: 8
  • Channels: 30
  • Learning rate: 1e-3
  • Time window size: 750

Quantitative classification performance

drow_table

Rquirements

python ≥ 3.8.10 numpy ≥ 1.20.3 pandas ≥ 1.2.5 scipy ≥ 1.6.2 torch ≥ 1.9.0

sklearn random os argparse collection

The predicted values for test are stored in files under each model name + "prediction".