Polygonal Shape-based Features for Pose Recognition using Kernel-SVM

Abstract

Pose recognition is an intriguing and challenging problem particularly in surveillance, inspection, etc. that lies in computer vision. This paper presents an efficient human walking and abnormal poses recognition system based on kernel-support vector machine (KSVM) using a novel feature set based on polygonal shape generalization on the human silhouette. The Shapiro-Wilk test was conducted to assess the data distribution and it summarized that the test rejected the hypothesis of normality for all features. Therefore, an inferential Mann-Whitney U test was performed to evaluate the proposed feature set statistically and results showed that all features were significantly different between the groups of poses (p < 0.001). Three kernel models: linear, polynomial and radial based function were adopted for SVM to classify the walking and abnormal poses. Results obtained showed that all three kernels of the KSVM classifiers performed well with accuracies of more than 95%. However, further experiments proved that the polynomial KSVM yields the best accuracy of 99.96%. Thus, it can be concluded that the proposed polygonal shape-based feature set is best paired with the polynomial KSVM for abnormal pose detection task.