T10-Gait

Foot segment identification from footsteps on pressure pads

Google Drive folder is here: https://drive.google.com/open?id=0B2WhXU_IjE68cjlmMzZZdV9iMm8

Proposal is here: https://drive.google.com/open?id=1vuIcVz8KHbXqyjY4PWqkA_2YfWaEg1bcKuxf7oa9vcg

Conditional Random Field training

File main.m trains a Conditional Random Field (CRF) model. It is heavily based off of sophisticated_example.m.

Setup:

1. Download JustinsGraphicalModelToolbox from the Google Drive, or from Slack. Unzip it and move it into the root of the project. The folder "JustinsGraphicalModelToolboxPublic" should be in the same folder as this README.
2. Rename JustinsGraphicalModelsToolboxPublic/External/toolbox/external/other/savefig.m to JustinsGraphicalModelsToolboxPublic/External/toolbox/external/other/savefig_m.m. We rename it because it has the same name as a MATLAB builtin and causes problems.
3. Add the graphical model toolbox to MATLAB's path, recursively. Do this by running >> addpath(genpath('JustinsGraphicalModelsToolboxPublic')); in MATLAB.

Train:

1. In MATLAB, run >> main.

Semi-Automated Foot Labelling

File labelling_automation/label_foot.m is a function which will guide you through converting an image of a labelled foot into proper labelled foot data which the system can consume in model training.

The paths are hardcoded. It must be run from within labelling_automation.

Example: labeling FinalData/NP40/3108

>> cd labelling_automation
>> label_foot('NP40', '3108'`)

In this case, the files FinalData/NP40/3108.jpg and FinalData/NP40/3108.lst must exist. The jpg file is an image of an expert-segmented image, and the lst file is the footstep data file for the given image.

The first step shows a picture of the labelled foot and instructs the user to resize a rectangle to fit the outline of the colored squares. Each step after that shows a picture of the labelled foot and instructs the user to trace one labelled region.

The region mask is then converted from pixel-coordinates to .lst-file coordinates and the result is written to a .mat file in labels_data.

It also moves the matching .lst file to training_test_data.

Files, Directories

Directory training_test_data contains pedobarograph (pressure pad) data.

• Each .lst file represents one reading of a single footstep on the pedobarograph. It is read using pedo_extract.m.

Directory labels_data contains labeled matrices.

• Each .mat file is a labelled matrix corresponding to the max-pressure map of the .lst file with the same id, eg pb_data/NP10_2651.lst corresponds to labels_data/NP10_2651.mat.
• The labelled matrix contains 1s for the great toe, 2s for the lateral forefoot, 3s for the medial forefoot, 4s for the heel, and 0s for all other entries.

File pedo_extract.m (Written by: Quinn Boser, July 2013) takes a lst file and returns

1. a 3d struct array with dimensions (x,y,time)
2. a 2d max-pressure map with dimensions (x,y)
3. the length of the x dimension
4. the length of the y dimension
5. the length of the time dimension

File pedo_format.txt describes the format of an .lst file. It is used by pedo_extract.m.

File get_params.m reads the number of rows, columns, and frames from an .lst file, where frames is the length of the time dimension. It is used by pedo_extract.m.

File JaccardScore.m takes 2 matrices, A and B, and returns their Jaccard Score.

File silly_example.m is a small example of training a Conditional Random Field (CRF) on noise. It is pulled from http://users.cecs.anu.edu.au/~jdomke/JGMT/#binarydenoising under the MIT License.

File sophisicated_example.m is an example of training a Conditional Random Field (CRF) on the Stanford Backgrounds dataset. It is pulled from http://users.cecs.anu.edu.au/~jdomke/JGMT/#backgrounds under the MIT License.