skeleton-HMM

An implementation of the paper "Skeleton-based abnormal gait detection" (Sensors, MDPI 2016)

Requirements

• Python
• Numpy
• Scipy
• Scikit-learn
• hmmlearn
• Matplotlib

Notice

• The code was implemented to directly work on DIRO gait dataset
• Please download the skeleton data and put the npz file into the folder dataset

Usage

python main.py -l 0 -w 5 -s 24 -o 43 -f 0

• -l: use leave-one-out cross-validation (boolean)
• -w: width of smoothing window (int)
• -s: number of HMM's states (int)
• -o: number of HMM's observations (int)
• -f: write results to file (boolean)

Example of output

Default training and test sets

test subject(s): [1 3 6 7]
Load normal gaits of 5 subjects for training...
processing normal skel. of subject 0
processing normal skel. of subject 2
processing normal skel. of subject 4
processing normal skel. of subject 5
processing normal skel. of subject 8
Load test data...
processing skel. of subject 1
processing skel. of subject 3
processing skel. of subject 6
processing skel. of subject 7
window width = 5, states = 24, observations = 43
kmeans dimension: 7
TEST RESULTS
Full sequence:   AUC = 0.898 --- EER = 0.250
Cycle:           AUC = 0.792 --- EER = 0.277

Leave-one-out cross-validation

...
Leave-one-out means
Full sequence:   AUC = 0.806 --- EER = 0.173
Cycle:           AUC = 0.607 --- EER = 0.417