Human Action Recognition from Variable Silhouette Energy Images

Abstract

Recognizing human actions is an important issue in the computer vision community. Human action recognition becomes more challenging when variability areas such as, anthropometric variation, phase variation, speed variation, camera view variaticn, individual variations in appearance and clothes of people, changes in light and view point and so on. In this thesis, we propose a spatio-temporal silhouette representation, called silhouette energy image (SEl) and silhouette history image (SF11), to characterize motion and shape properties for recognition of humau movements such as, human actions, activities in daily life. We also proposed variable silhouette energy image for different variable situations. To address the variability in the recognition of human actions several parameters such as, anthropometry of person, phase (starting and ending state of action) speeds of the actions, camera observation (distance from camera, tilting motion and rotation of human body) and view variations are proposed. The SEI and Sill are constructed using the silhouette image sequence ofan action. The span or difference of the end time start time is used to make SF11. We extract the features based on geometrical shape moments. Using the features, we generate a unified description of model by learning (lie multi-class SVM for each action. Finally we recognize action using action model for any arbitrary image sequence. We tested our approach successfully in the indoor and outdoor environment. Our experimental results show that the proposed method is robust, flexible and efficient.