Hierarchical structure for Modeling Human actions and Multi-Classifier Approach

Abstract

-Automatic action recognitionis one of the challenging tasks in recent research. Complex human actions are composed of simpler motion patterns. This research relates to give a new direction for identifying human actions using hierarchical structure of motion patterns. The main contribution of our research is that we proposed hierarchical classifier based action recognition system. In past years, no analysis is done for recognition using different classifiers for each complex human actions with exploration of simpler motion patterns. We modeled our system by the fact that each human action originates from basic human movements. Human actions such as running and walking originate from legs movement and hand waving and hand clapping originate from hands movement. First different classifiers were used in our proposed single classifier hierarchical structure and then best performed classifier for each action is selected and applied in multi-classifier hierarchical structure. For classifier inputs, Spatio-Temporal Interest Points (STIP) are extracted using SIFT features from 50 consecutive frames of each action. Covariance of STIP features among action frames are used as feature vector for classification using KNN, SVM and Naïve Bayes. Hierarchical structure is implemented using single classifier approach where each classifier is used separately at each level of hierarchy. Analysis is done and it is concluded that each classifier behavior and performance is different for each action. Best classifiers are selected and integrated in hierarchical structure using multi-classifier approach. Results show that multi-classifier approach in hierarchical structure has improved results as compared to single classifier approach.