Multiple feature representations from multi-layer geometric shape for hand gesture analysis

Research on hand gesture recognition has been intensified in the last decade. There has been successful work on gesture recognition by using shape and texture features from 2D images and videos. However, it is still a challenging task in handling various conditions with hand scale variations, hand rotations, and hand ambiguity due to finger occlusions. Recent work has started using depth images for such study. However, the low quality of the depth data increases the difficulty for geometric feature extraction and representation. In this paper, we analyze multiple types of features of hand gestures (or hand postures), and propose a novel scale-invariant representation which encodes topological features and shape features both locally and globally. Our approach is applicable to depth images as well as to regular 2D images. We evaluate the new descriptor in terms of rotation robustness, person independence, and modality independence, and further compare it to the existing four state-of-art feature representations. The results demonstrate the robustness and reliability of our proposed descriptor in recognizing sixteen types of hand gestures.