Efficient High Cross-User Recognition Rate Ultrasonic Hand Gesture Recognition System

Hand gesture recognition is the key technique to realize free air hand gesture human-computer interaction, which is promising and attractive for a wide range of emerging ubiquitous applications. Ultrasonic-based solutions are lighting-condition independent and low-power and have little interference. A low-complexity 40 kHz ultrasonic-based hand gesture recognition system featuring a novel classification algorithm robust to temporal variations of motions is proposed in this paper. It employs one transmitter, three receivers, and highly digitalized front-end hardware. A coherent pulse train ultrasonic system with moving target detection capability is built to collect the RDM (range Doppler map) streams of gestures. During training and inference, the classifier algorithm aligns the sequential features extracted from the RDM of each frame with the template sequence of each class by dynamic time warping in advance. For each class, a two-class random forest that determines whether the testing sample belongs to the class according to the aligned features is trained. Then, the prediction is made by comparing the probabilities of the testing sample belonging to different classes. Experiments show that the proposed classifier trained based on 6 people has a better leave-one-subject-out validation accuracy than the competitors. It can identify 8 3D gestures with 93.9% accuracy in 37 ms on a PC. Its model size is 700 Kbytes. It is a better choice for the user-independent embedded applications.