Estimating signal-dependent noise (SDN)-based motion variations to enhance gesture recognition

Abstract

Gesture recognition is an important manoeuvre that uses known data patterns for comparison and/or training the recognition algorithm. Previously learned gestures can be recognized irrespective of the environment and position. However, even if the same gesture is repeated, variations are inevitably generated because of unconscious muscle movements. Such variations in motion can adversely affect the accuracy of gesture recognition algorithms. Extensive training data is conventionally used to reduce this effect. In this paper, we propose predicting such variations in gesture motions from a single training data-set. We first generate muscle activation data from the measured gesture trajectory using musculoskeletal simulation. Signal-dependent noise is then superimposed on the muscle activation data to generate 100 predictions containing noise variations. Back-calculation is performed on the predicted muscle activation. The final output is acceleration data in the X and Y axes, which are segmented to calculate the segment-wise variation. The inverse of variation is named as reliability unique to each segment. A 26% increase in the F-measure was observed when we compared the best-case scenarios of the proposed algorithm with conventional dynamic time warping gesture recognition without segmentation and consideration of motion variation.