A discrete Hidden Markov models recognition module for temporal series: Application to real-time 3D hand gestures

Abstract

Abstract-This work studies, implements and evaluates a gestures recognition module based on discrete Hidden Markov Models. The module is implemented on Matlab and used from Virtools. It can be used with different inputs therefore serves different recognition purposes. We focus on the 3D positions, our devices common information, as inputs for gesture recognition. Experiments are realized with an infra-red tracked flystick. Finally, the recognition rate is more than 90% with a personalized learning base. Otherwise, the results are beyond 70%, for an evaluation of 8 users on a real time mini-game. The rates are basically 80% for simple gestures and 60% for complex ones.

Recognition of occupational therapy exercises and detection of compensation mistakes for Cerebral Palsy

“…Because of this, research that compares different algorithms using the same set of gestures is inspected thoroughly. There are effective HMM-based gesture recognition solutions for Kinect time series joint data [24,[34][35][36]. HMMs are also extensively studied in sign language recognition, which is similar to exercise recognition in principle.…”

“…Specifying a constant threshold value does not work because the likelihoods of the models fluctuate altogether depending on the input properties, the length of the observation sequence being one of them [36,55]. Therefore, a mechanism should be devised that would produce an adaptive threshold value.…”