D2
Computer Vision and Machine Learning

Educational Activity Recognition Framework and Dataset

The framework is implemented in MATLAB and has the below features. The framework is accompanied with an educational dataset as well as a number of example scripts that produce all the results described in the paper. The framework and dataset can be downloaded HERE.

General

  • Support for arbitrary number and type of sensors with different configurations and placements, participants, as well as activity classes
  • Fully parametrized and customizable (window size, step size, number of features, etc)
  • Different evaluation methods (frame-based, activity event-based)
  • Plotting functions (confusion matrix, feature ranking, feature distribu- tions, PR- and ROC-curves, parameter sweeps)
  • Easily extendible with new classifiers, features, etc.
  • Large number of sample evaluations (all of those described in the paper)
  • Experiments-Pipeline: Effortless adding of new experimental settings

Classifiers

  • Discriminant Analysis (Matlab)
  • Discrete/continuous Hidden Markov models (HMMToolbox, Kevin Murphy)
  • Support vector machine (SVM light/liblinear, C++)
  • JointBoosting (Torralba, Implementation Wojek C++)
  • k-nearest neighbour (kNN, Matlab)
  • Nearest class centre (NCC, Matlab)
  • Naive Bayes (Matlab)

Segmentation methods

  • Sliding Window
  • Energy-based

Supported features

  • Mean
  • Variance
  • MCR (mean crossing rate)
  • ZCR (zero crossing rate)
  • Several FFT features

Feature selection methods

  • Minimum redundancy maximum relevance (mRMR, filter)
  • Sequential forward selection (SFS, wrapper)
  • Sequential backward selection (SBS, wrapper)

Decision fusion approaches

  • Early (feature-level)
  • Late (classifier level)

Evaluations

  • Person-independent
  • Person-dependent
  • Cross validation leave-one-person,fold-out
  • Evaluation tools:
    • Confusion matrix
    • Timeframe-based evaluation
    • Precision-recall curves + AUC, equal error rate (EER), average precision (AVGPREC)
    • Event based criterion (hit and min50)
    • Output label statistics