Gait Abnormality Detection And Classification

OS node that uses hidden markov models to classify a human's gait phase as swing or stance based on IMU data, foot pressure sensors, infrared and proximity data requires ROS indigo and the pomegranate library. Included are the data for 3 subjects, with 5 intervals per subject where:

Interval 1(and 2 for subject1 (not used)): normal walking
Interval 2: low foot clearance
Interval 3: small step width
Interval 4: excessive trunk sway
Interval 5: all of the previous combined in a random manner

(Code not documented is deprecated/not used)

Nodes

Record data

./launch/record.sh \<output_bagfile_name\> \<duration_in_minutes\>

Convert rosbags to useable form saving data, normalized data and timestamps in .matfiles:

Run:

For Single subject:

  roslaunch bag2mat.launch prefix:=\<path_to_folder/subject_name_recordingnumber\> file=/<path_to_parameter_file/>

The parameter file contains the topics for each imu, see /scripts/new_bags/sub1.yaml

For all subejcts:

  ./launch/bag2matnewbags.sh

This will create a number of .mat files for each input file containing each component recorded.

Annotation:

To annotate the data file with mocap data is required in a format similar to the files found in /scripts/<subject_number>.zip, that are in the follwoing format:

COMmovement: X-anterior-posterior trunk sway; Y-medio-lateral trunk sway.

Annotate a single subject:

rosrun gait_hmm_ros auto_annotate.py _prefix:=new_bags/subject1_1 _matfile:=new_bags/subject_one/subject01_normal

Annotate every subject

./launch/annotatenewbags.sh

This will create a file named like your input file with the addition of "_labels" in the end. It will contain the gait phase label for every timestamp created in the previous cript.

Data set creation:

Create different sets of data containing or excluding the sensors you want to use, based on /scripts/new_bags/<subject>_full_data.mat created with the previous scripts.

./batch_create_combos_\<sub1,sub2,sub3,gyro\>

Uses the /launch/create_data.launch script which accepts the following parameters:

batch_train: 0/1 whether to train classifier in batches or with single inputs
batch_test: 0/1 whether to test classifier in batches or with single inputs
use_quat: 0/1 whether to use the quaternion data when training the classifier
use_gyro: 0/1 whether to use the gyroscope data when training the classifier
use_accel: 0/1 whether to use the accelerometer data when training the classifier
use_com: 0/1 whether to use the compass data when training the classifier
use_ir: 0/1 whether to use the IR data when training the classifier
use_prox: 0/1 whether to use the proximity data when training the classifier
use_fsr: 0/1 whether to use the FSR data when training the classifier
rf: topic for foot IMU, empty to not use
rl: topic for lower leg IMU, empty to not use
rul: topic for upper leg IMU, empty to not use
m: topic for lower back IMU, empty to not use

Train classifiers:

Train Single Classifier

rosrun gait_hmm_ros create_single_classifier.py _folds:=\<number of training folds\> _norm:=\<1 or 0, use normalized data or not\>
_datafile:=\<first component of datafile to use se /launch/create_single_classifiers_1.sh for the format\>

Train Multpile Classifiers:

./launch/create_single_classifiers_(1-4).sh

The classifiers will be saved as pickled python objects in /scripts/trained_classifiers/

Receives input from IMUs and other devices and uses an ANFIS to classify the type of gait of the user (normal gait, excessive trunk sway, small step width, low foot clearance). Requires Matlab's Fuzzy Logic Toolbox, the 3Space IMU python API, numpy, and scipy. Tested in ROS Indigo

Sample bag files are located in scripts/new_bags/subject[number]_[trial].bag

Launch files

record.sh: waits 5 seconds and the starts recording available IMU frames
```
sh record.sh [filename] [duration-minutes]
```
batch_create_combos_x.sh: create different datasets using subsets of the available measurements
create_single_classifier_x.sh: create different classifiers using subsets of the available measurements
align.launch: used to align mocap files with data from rosbags
```
roslaunch align.launch prefix:=[datafile] mocap:=[mocap_file]
```
bag2mat.launch: used to create new matfiles from rosbags
```
roslaunch bag2mat.launch param_file:=[param file] prefix:=[file name]
```
param files contain the topic names for the different imus as "upper_leg_topic lower_leg_topic foot_topic"

auto_annotate.launch: annotates a matfile using a given mocap file

roslaunch auto_annotate.launch prefix:=[data file] prefix:=[mocap file]

create_data.launch: creates training files using subsets of the available sensor data, edit the launch file to change, 1 means uses, 0 means don't use. Creates normalized and non-normalized datasets and prints out their results in /new_bags. Data is saved as .mat files or pickles. Also save are the label for each reading, and trunk sway, step length and width for mocap data.

A complete data file has the following format:

[upper_leg_quaternion_x][upper_leg_quaternion_y][upper_leg_quaternion_z][upper_leg_quaternion_w][upper_leg_accelerometer_x] [upper_leg_accelerometer_y][upper_leg_accelerometer_z][upper_leg_gyroscope_x][upper_leg_gyroscope_y][upper_leg_gyroscope_z] [upper_leg_magnetometer_x][upper_leg_magnetometer_y][upper_leg_magnetometer_z] [lower_leg_quaternion_x][lower_leg_quaternion_y][lower_leg_quaternion_z][lower_leg_quaternion_w][lower_leg_accelerometer_x] [lower_leg_accelerometer_y][lower_leg_accelerometer_z][lower_leg_gyroscope_x][lower_leg_gyroscope_y][lower_leg_gyroscope_z] [lower_leg_magnetometer_x][lower_leg_magnetometer_y][lower_leg_magnetometer_z] [foot_quaternion_x][foot_quaternion_y][foot_quaternion_z][foot_quaternion_w][foot_accelerometer_x] [foot_accelerometer_y][foot_accelerometer_z][foot_gyroscope_x][foot_gyroscope_y][foot_gyroscope_z] [foot_magnetometer_x][foot_magnetometer_y][foot_magnetometer_z] [FSR reading][IR reading][Proximity Reading]

ROS Scripts

align_mocap.py: Used to align mocap data and rosbag recordings, use with align.launch
arduino_listener.py: Used to read arduino data and republish it as a ROS message
auto_annotate.py: Annotates a matfile using a mocap output file (see auto_annotate.launch)
fsr_annotate_2.py: annotates data using just the FSR input
create_data_combos.py: creates matfiles using subset of the available measurements (see create_data.launch)

Matlab Scripts

bigOlClassifierScript2: creates and evaluates all possible classifiers
create_anfis_list: reads available files in ~/full_data and creates a list of classifiers to test
build_anfis_classifier_loop.m: for every classifier in the list created with create_anfis_list, build and evaluate the classifier
evaluate_anfis_classifier_loop.m: just evaluates the classifier and saves the results
create_and_evaluate_classifier.m: creates a classifier based on a given workspace, evaluates the classifier and saves the results
create_anfis_classifier.m: splits data into test and training data, builds a classifier and saves the workspace
plots_and_stats.m: loads a built classifier, evaluates it, and creates a graph of its performance

Unused/Deprecated

HMM classifiers, can be used for reference

create_single_classifier.py
create_distributed_classifier.py
dis_class_2.py
distributed_classifier_2.py
test_classifiers.py
test_classifiers.distributed.py
create_anfis_classifier_phase.m
create_anfis_classifier_normal.m

Random

fsr_annotate.py: see fsr_annotate_2.py

AndLydakis/gait_hmm_ros