Thesis (Ph.D.)--Chulalongkorn University, 2012

A five postures recognition algorithm from a human video was proposed. To make the algorithm practical to real surveillance applications, several contexts concerning clothing, camera factors, sequence of acting postures, and subject's body size are not neither constrained nor controlled. A new set of features based on only the geometrical information of stationary and temporal posture envelops was introduced. Our algorithm can handle both stationary and moving postures, which are walking, standing, lying, sitting, and bending. The algorithm was tested in the following two aspects such as the efficacy of proposed features, comparison with Juang, C.F. and Chang, C.M.’s research (2007) using Neural Fuzzy Network classifier in four (excluding walking) and five postures classification, respectively, and applicability to various data sets, testing the robustness of algorithm in various data sets. Four video clips from Weizmann human posture data set, UIUC complex active data set, CMU graphic lab motion database, and our own data set were experimented. With respect to our own data sets our average accuracy achieved 93.27%. But when our algorithm was tested will other data sets, our accuracy is higher than 95%.