featuregait

Feature Learning for Accelerometer based Gait Recognition

Code repository of paper: Feature Learning for Accelerometer based Gait Recognition, submitted to ICANN 2020

Used datasets

• [ZJU-GaitAcc] - http://www.ytzhang.net/datasets/zju-gaitacc/

  • session 0 - 22 subjects
  • session 1 - 153 subjects
  • session2 - 153 subjects

• [IDNet] - http://signet.dei.unipd.it/research/human-sensing/

  • 50 subjects with various number of sessions

Segmentation

• FRAME-based: length = 128 samples

• CYCLE-based:

  • Cycle length statistics ZJU-GaitAcc: https://github.com/margitantal68/featuregait/blob/master/statistics/ZJUGaitAccel.png
  • Based on statistics, gait cycles were normalized to 128 samples

Features

• RAW - use raw accelerometer data as features - 3 x 128 = 384 (ax - ay - az)
• HANDCRAFTED - 59 ad-hoc statistical features (For details see: https://github.com/nemesszili/gaitgmm)
• UNSUPERVISED feature extraction - autoencoders
  • DENSE autoencoder
  • Fully Convolutional (FCN) autoencoder
  • Time Convolutional (TimeCNN) autoencoder

Identification/Classification - based on a single gait segment (FRAME or CYCLE)

• Random Forest - 100 trees
• Two protocols:
  • SAME-DAY: using data from a single session - 10-fold CV - evaluated for session 1 and 2 separately (153 subjects)
  • CROSS-DAY: training - session 1, testing - session 2

COMPARED to SUPERVISED feature extraction - end-to-end deep models (FCN, ResNet)

• Two protocols:
  • SAME-DAY: using data from session 1 (train-validation-test: 60%-20%-20%)
  • CROSS-DAY: training - session 1, testing - session 2

Code

The code is divided as follows:

• The rawfeatures.py contains the code for creating raw features

• The handcrafted.py contains the code for extracting handcrafted features

• The main.py python file contains the necessary code to run an experiment.

• The util folder contains the following:

  • const.py - constants and enumerations
  • identification.py - identifications using various features
  • load_data.py - various data loading and reshaping
  • myplots.py - plots
  • settings.py - different configurations for running an experiment
  • statistics.py - cycle length statistics for ZJU-GaitAcc

• The autoencoder folder contains the following:

  • autoencoder_common.py - common functions used by different types of autoencoders
  • autoencoder_dense.py - Dense (MLP) autoencoder model
  • autoencoder_cnn.py - CNN autoencoder models (FCN, TimeCNN)

• The notebooks folder contains two Jupyter notebooks:

  • ZJU cycles.ipynb - end-to-end FCN and ResNet models - cycle based measurements
  • ZJU frames.ipynb - end-to-end FCN and ResNet models - frame based measurements