Gesture recognition based on Hidden Markov Model from sparse representative observations

Hand gesture recognition plays an important role in human computer interaction, virtual reality, and so on. In this paper, we focus on how to generate efficient observations after feature extraction in Hidden Markov model (HMM). Vector quantization such as kmeans clustering algorithm is usually applied to generate codebooks in HMM-based methods. Unlike traditional vector quantization, we use sparse coding (SC) and HMM to achieve the task of hand gesture recognition, which we call ScHMM. Sparse coding provides a class of algorithms for finding succinct representations of stimuli. In the training stage, feature-sign search algorithm and Lagrange dual are applied to obtain codebook and in the testing stage, feature-sign algorithm is used to get efficient observations. We evaluated our method on public database. ScHMM compares favorably to state-of-the-art methods, namely HMM, conditional random fields, hidden conditional random fields and latent dynamic conditional random fields.