Semantic Recognition of Human-Object Interactions via Gaussian-Based Elliptical Modeling and Pixel-Level Labeling