Multi-class support vector machine for human posture classification using a simplified shock graph

• Published in: 2008 International Symposium on Information Theory and its Applications, ISITA2008
• Main Author: Shahbudin S.; Hussain A.; Tahir N.M.; Samad S.A.
• Format: Conference paper
• Language: English
• Published: 2008
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-77951105231&doi=10.1109%2fISITA.2008.4895452&partnerID=40&md5=a4400143bec6069cb1ce676b8d054eb4

In this paper, we intend to present and discuss the validity and effectiveness of shock graph as feature vectors for human posture classification task. Basically, a shock graph is a shape abstraction that decomposed a shape into a set of hierarchically organized primitive parts. The shock graph that represents the silhouette of an object in terms of a set of qualitatively defined parts, organized in a hierarchical, directed acyclic graph, is used as a powerful representation of human shape in our work. To be exact, the paper aims to highlight the efficacy of the simplified shock graph (SSG) derived as a result of pruning from the original shock graph as revealed in [7]. We believe our SSG provides a compact, unique and simple way of representing human shape. In doing so, we have chosen to test the feature vectors of the SSG with three different multi-class Support Vector Machines (SVM) classifier models. In this study, there are three different method of SVM classifier will be used, namely the one-against-all (OAA), one-against-one (OAO) and the bias support vector machines (BSVM). In addition, Gaussian radial basis function (RBF) was opted as kernel for all three techniques with the Sequential Minimal Optimization (SMO) solver applied for training SVM of the two-class problem for both OAA and OAO methods along with Nearest Point Algorithm (NPA) solver used to train the multi class BSVM. The training time, regularization parameter, C values together with other parameters such as the number of support vectors, testing time and classification rate will be evaluated. Initial results demonstrated that perfect classification rate is attained for all three SVM models with the BSVM approach contributed the least number of support vectors. The best performance for training and testing time is owned by the OAO method. As such, the validity of SSG as feature vectors to represent human postures is confirmed based on the findings from all three methods of SVM applied.