A Quantitative Performance Assessment of surface EMG and Accelerometer in Sign Language Recognition

Sign language predominantly makes use of different hand postures and hand motions. The combined use of surface electromyogram (sEMG) and accelerometer sensors in a wearable sign language recognition system has been shown to enhance performance as compared to either of the sensing modalities. However, since sEMG is prone to motion artefacts, noises arising due to sweat on skin surface besides being costlier as compared to accelerometers, there has been a discussion in literature regarding the development of sign language recognition system with only motion sensors. In this work, the utilities of sEMG and accelerometer are analysed to reveal the scenario under which each of these modalities contribute the most. For a set of 1200 recordings, the overall accuracy when only sEMG or accelerometer signals are used is found to be 86.3% and 82.1% respectively, whereas their combined use yields an accuracy of 87.5%. Although, the inclusion of accelerometer data with the sEMG signals in a sign language recognition system is found to improve the overall recognition accuracy of the signs, it is demonstrated that under certain conditions the accelerometer data does not contribute much towards sign recognition. In fact, inclusion of accelerometer data with the sEMG signals under these conditions may adversely affect the classification accuracy of the sign. However, for another category of signs, accelerometers are sufficient for classification and sEMG is not required. The supporting results are also tested for significance using statistical analysis.