HUMAN ACTION CLASSIFICATION USING MOTION AND APPEARANCE FEATURES FOR ACTIVITY UNDERSTANDING AND ANOMALY DETECTION IN VISUAL SURVEILLANCE

Abstract

Human action recognition is one of the interesting research areas in computer vision. It is an important component of the automated surveillance system which is needed to reduce the insufficiency and inefficiency of human’s role in the system. However, one of the challenges in action recognition is the extracting appropriate features for classification. Good features should contain both motion and appearance information of human. Also, the feature extraction process should automatically adapt to the speed variation of human actions when applying the system to the different performers, actions, and datasets. In this thesis, we propose the Adaptive Key Frame Interval (AKFI) feature extraction to segment human action into primitive action subsequences. The interval length is automatically changed based on the action characteristic and speed of the performer. Once key frames are detected, the features within a segmented period are encoded by Adaptive Motion History Image (AMHI) and Key Pose History Image (KPHI). The features contain both appearance and motion information of human actions. The experimental results demonstrate that the system can differentiate the unusual action from the normal situation. Also, AMHI and KPHI can effectively classify action compared to other well-known algorithms.