3D hand gesture recognition from one example

Abstract

In a typical recognition system, the inclusion of more training data is likely to increase the recognition rate. However, it is not easy to obtain large training sets. Focusing on practical applicability such as controlling home appliances, we propose a hand gesture recognition method from one example that is computationally efficient and can be easily implemented. 3D hand motion trajectory is achieved from a depth camera and then normalized for translation invariant feature extraction. Based on the simple K-NN classifier, we develop a pattern matching method by combining the DTW (Dynamic Time Warping) algorithm and a statistical measure for similarity between two random vectors. We conducted computational experiments on hand gesture data and compared the results with those derived via conventional DTW recognition.