Modified CRF algorithm for dynamic hand gesture recognition

In this paper, a modified CRF algorithm is proposed for recognition of vision-based dynamic hand gestures. This algorithm abandons the condition necessary for Hidden Markov Models that the action sequences must be independent. And dynamic hand gestures are classified by some most representative segments (MRSs) rather than the full gestures themselves. First, the Longest Common Sequence (LCS) is employed to extract the most representative segments from dynamic gestures which are then used to train Conditional Random Fields (CRF). In a recognition stage, MRS of the unclassified trajectory is sent to CRF. Experiment results show that this algorithm (defined as MRS-CRF) has significant advantages over HMMs in accuracy and CRF itself in simplification.