HUMAN ACTION RECOGNITION USING MONOTONIC TRIANGULAR CONTEXT BASED SHAPE FEATURES

Recognition of human action from video sequences is an active area of research in computer vision. In this paper, we present a novel shape descriptor to represent the boundary of human silhouette using monotonic triangulation technique. The proposed shape descriptor best captures the orientation information and extracts two important features namely, Triangulated Shape Orientation Context (TSOC) and Centroid Orientation Context (COC). This approach is compact, view-invariant and independent of clothing conditions for the number of frames which represents human action. After background subtraction, we extract the proposed features and a specific discrete Hidden Markov Model (dHMM) is trained for each action, grouping the sptio-temporal manifolds. We tested the robustness of our approach using Inria Xmas Motion Acquisition Sequences (IXMAS) and Virtual Human Action Silhouette (ViHASi) datasets. We also demonstrated the performance using real-world scenes to emphasize the potential usefulness in practice.