Segment motion using a 2-state Gaussian HMM

segment_motion.py

"""
Segment an acceleration or gyroscopic CSV file into motion/non-motion segments
using a 2-state HMM and Savitzky–Golay filter as preprocessing
"""
import sys
import pandas as pd
import matplotlib.pyplot as plt
from hmmlearn.hmm import GaussianHMM
from scipy.signal import savgol_filter

WINDOW_LENGTH = 15  # Must be odd
POLY_DEGREE = 2


def segment(x, y):
    # use a 2-state HMM to separate motion/non-motion states
    hmm = GaussianHMM(n_components=2)
    hmm.fit([y])
    states = hmm.predict(y)
    # the motion state is the one with the max mean, assuming y is some type of
    # measure of motion (acceleration, velocity, gyroscope, etc.)
    motion_state = hmm.means_.argmax()
    return pd.Series(states == motion_state)


if __name__ == '__main__':
    if len(sys.argv) != 3:
        print('Usage: $ python segment_motion.py [input] [output]')
        sys.exit(1)

    df = pd.read_csv(sys.argv[1])

    # mean absolute acceleration from 3 axes
    y = df[['X_value', 'Y_value', 'Z_value']].abs().mean(axis=1)

    # play the filter params or try using a different filter
    y = savgol_filter(y, WINDOW_LENGTH, POLY_DEGREE)

    # motion is a Series of booleans, True where there is motion
    motion = segment(df['time'], y[:, None])

    lines = plt.plot(df['time'], df[['X_value', 'Y_value', 'Z_value']])

    # gray spans indicate motion
    for _, g in df['time'][motion].groupby(motion.diff()[motion].cumsum()):
        plt.axvspan(g.iloc[0], g.iloc[-1], facecolor='k', alpha=0.2, linewidth=0)

    plt.legend(lines, ('X', 'Y', 'Z'))
    plt.show()

    motion.to_csv(sys.argv[2])

Gray areas indicate detected motion.

Example using accelerometer:

Example using gyroscope: