Combining HMM and SPSM for sign language recognition

The research of sign language recognition has great academic value and broad application prospect.In recent works on sign language recognition,Hidden Markov Models(HMMs) has played an important role.But the framework of HMMs makes an assumption that the observations in the same state are independent and identically distributed,which is not consistent with sign language signals sometimes.Inspired by Polynomial Segment Models(PSMs) that can model the consecutive frame correlation well,in this paper we propose a type of simplified PSMs,in which Mahalanobis distance is used as the similarity metric.Experimental results show that after combining conventional HMMs and the simplified PSMs with summation of normalized posteriori,the average relative accuracy can be improved by 13.38%.As result,our method is superior to conventional HMMs.