Cross-Scene Sign Language Gesture Recognition Based on Frequency-Modulated Continuous Wave Radar

Abstract

This paper uses millimeter-wave radar to recognize gestures in four different scene domains. The four scene domains are the experimental environment, the experimental location, the experimental direction, and the experimental personnel. The experiments are carried out in four scene domains, using part of the data of a scene domain as the training set for training. The remaining data is used as a validation set to validate the training results. Furthermore, the gesture recognition results of known scenes can be extended to unknown stages after obtaining the original gesture data in different scene domains. Then, three kinds of hand gesture features independent of the scene domain are extracted: range-time spectrum, range-doppler spectrum, and range-angle spectrum. Then, they are fused to represent a complete and comprehensive gesture action. Then, the gesture is trained and recognized using the three-dimensional convolutional neural network (CNN) model. Experimental results show that the three-dimensional CNN can fuse different gesture feature sets. The average recognition rate of the fused gesture features in the same scene domain is 87%, and the average recognition rate in the unknown scene domain is 83.1%, which verifies the feasibility of gesture recognition across scene domains.