activity-recognition-KTH

Human Activity Recognition based on Summarized Semi-detailed Frame Information and Contextual Features

This repository contains code for our paper:

D. Chatterjee, C. Arora, S. Chakraborty and S. K. Saha, "Human Activity Recognition based on Summarized 
Semi-detailed Frame Information and Contextual Features," 2020 IEEE Calcutta Conference (CALCON), 
Kolkata, India, 2020

Please considering citing our paper if you find our code useful:

@inproceedings{chatterjee2020human,
  title={Human Activity Recognition based on Summarized Semi-detailed Frame Information and Contextual Features},
  author={Chatterjee, Dibyadip and Arora, Charu and Chakraborty, Saurajit and Saha, Sanjoy Kumar},
  booktitle={2020 IEEE Calcutta Conference (CALCON)},
  pages={127--131},
  year={2020},
  organization={IEEE}
}

Requirements

• Python3
• Numpy
• Matplotlib
• OpenCV
• sklearn

File Structure

Within ./src/main.py, following python files are available:

• main.py The main file which will be run on the terminal with command line arguments for different functionalities.
• motion_2x2.py This file contains the function that produces motion based features (HOF) when the POI is divided into 2x2 grids.
• motion_3x3.py This file contains the function that produces motion based features (HOF) when the POI is divided into 3x3 grids.
• motion_4x4.py This file contains the function that produces motion based features (HOF) when the POI is divided into 4x4 grids.
• contextual_2x2.py This file contains the function that produces motion based contextual features when the POI is divided into 2x2 grids.
• contextual_3x3.py This file contains the function that produces motion based contextual features when the POI is divided into 3x3 grids.
• contextual_4x4.py This file contains the function that produces motion based contextual features when the POI is divided into 4x4 grids.
• orientation_features.py This file contains the functions that produce orientation based (HOG) normal and contextual features i.e. features A and C as referred in the paper respectively.
• classifier.py This file contains the functions for training, testing and preparing data for our baseline classifier, Support Vector Machine (SVM).
• utils.py This file contains utility functions required for proper functioning of the other files.

Working

The 'cascadG.xml' can be downloaded from here.

Processing features

python3 main.py process

For e.g. if you want to process all the contextual features (described in the paper as C+D) the "process" section of the ./src/main.py is supposed to look like

#a = motion_2x2(str)
#b = motion_3x3(str)
#c = motion_4x4(str)
#d = HOG_central(str)
# an added context feature is also processed for motion to get consistency between the valid videos processed
#e, f_n = contextual_2x2(str)

a, f_na = contextual_2x2(str)
b, f_nb = contextual_3x3(str)
c, f_nc = contextual_4x4(str)
d = HOG_context(str)

b = np.concatenate((a,b,c,d), axis=0)

The processed features get saved in the ./src/bin folder as a .npy file.

Training individual features

Load the required numpy feature file from the ./src/bin folder and run the following command on the terminal

python3 main.py train

Before training, set the SVM parameters through grid search for which change the code under "train" section in the ./src/main.py as

# for training and classification
#SVM(feature_vector)

# for grid search
SVM_grid(feature_vector)

Training an ensemble of classifiers

Load any 2 required numpy feature files from the ./src/bin folder and run the following command on the terminal

python3 main.py train_ensemble

Create an ensemble of features (concatenation) from any two individual features

Load any 2 required numpy feature files from the ./src/bin folder and run the following command on the terminal

python3 main.py concat_features

Experimental Results

Results are measured on KTH Dataset. It consists of 599 videos which belong to 6 different classes viz. boxing, handclapping, handwaving, jogging, running and walking.

Features	Accuracy
Orientation Based (A)	66.39%
Motion Based (B)	88.23%
Orientation Based Contextual Feature (C)	73.87%
Motion Based Contextual Feature (D)	84.87%
A+B	89.07%
C+D	87.34
(A+B)*(C+D)	92.00%

where

• ' + ' : Ensemble of features (Concatenation)
• ' * ' : Ensemble of classifiers

License

MIT

Author

Dibyadip Chatterjee