Dynamic Hierarchial Clustering Based Human Action Classification

Abstract

Automatic visual surveillance systems could play an important role in supporting and eventually replacing human observers. The proposed frame work contains four modules viz detection, tracking, feature extraction and action classification. The person has to be found in the image plane first, using robust background subtraction by modifying the parameters in Gaussian Mixture Model (GMM), assuming a static camera. Centroid based tracking is done consecutively. Silhouette shaper based features are extracted, including aspect ratio of minimal bounding box of human silhoutte, approximated eccentricity, normalized central moments. Human actions are classified by grouping, enhancing efficient hierarchial clustering. In this paper a multi-sample-based similarity measure has been proposed where HMM training and distance measuring are based on multiple samples. These multiple training data are acquired by a dynamic hierarchical clustering (DHC) method. By iteratively reclassifying and retraining the data groups at different clustering levels, the initial training and clustering errors due to overfitting will be sequentially corrected in later steps. Human behaviours are obtained from clustered data set and trained neural network kit to describe their appropriate classes. The results are evaluated by means of sensitivity, specificity and accuracy.