Human-Activity-Recognition

This project classifies different human activities into their respective actions using the LibSVM library. The data is taken from microsoft kinect dataset which can also be downloaded from here. The data is converted in to three different representations:

• Relative angles and distances (RAD)
• Histogram of joint position differences (HJPD)
• Histogram of oriented displacements (HOD)

Dependencies

• python 2.7
• Numpy
• CSV
• pandas
• Libsvm

Run Instructions

git clone
python rad.py
python rad_d2_train_2.py
python accuracy.py 

Contents

• The train and test folders contain the training and testing data which can also be downloaded from here

• The files to generate the skeleton based representation models are Rad.py , HJPD.py, HOD.py

• The files to generate the Libsvm format training and testing data are Rad_d2_train_2.py , hjpd_d2_train.py, HODfull_train.py

• The file to predict the accuracy of the training and testing data are accuracy.py

• The repository contains 3 folders rad , hjpd , hod. Each folder contains the representation files (eg: rad_d2, rad_d2.t, its grid search graphs files, and the prediction file eg: pr1 )

RAD: prediction file-> pr1,pr2,pr3

HJPD: prediction file-> phj1 , phj2

HOD: prediction file-> ph1

Results

• The best values for C and gama are:

• RAD: C = 2 gama= 0.125

• HJPD: C = 2 gama= 0.00781

• HOD: C = 128 gama = 0.00195

Accuracy for the different representations.

• RAD = 54.16%

• HJPD = 83.333%

• HOD = 53.09%