Gesture Recognition Method Using Acoustic Sensing on Usual Garment
Gesture Recognition Method Using Acoustic Sensing on Usual Garment
复制标题
DOI：
10.1145/3534579
复制
发表时间：
2022-07
影响因子：
--
通讯作者：
Takashi Amesaka;Hiroki Watanabe;M. Sugimoto;B. Shizuki
中科院分区：
文献类型：
--
作者：
Takashi Amesaka;Hiroki Watanabe;M. Sugimoto;B. Shizuki
在这项研究中，我们展示了一种新的手势识别方法，基于服装的手势输入方法，使用主动和被动声学传感。我们的系统包括一个压电扬声器和麦克风。扬声器发射超声波扫频正弦信号，麦克风同时记录通过服装传播的超声波信号和服装上的手势产生的摩擦声。我们的方法识别各种手势，如捏，扭，触摸和滑动，通过结合主动和被动的声学传感。我们的方法的一个重要特征是，它不需要嵌入专用的服装或刺绣，因为我们的系统只需要一对压电元件用磁体附接到通常的服装。我们进行了识别实验的11个手势的前臂与四种类型的服装由不同的材料和识别实验的五个单手手势的衬衫的按钮和裤子的口袋。每用户分类器的结果证实，对于具有四种不同类型的服装的11个手势和假设实际使用而选择的5个手势，f分数分别为83.9%和95.9%。此外，我们还证实了该系统可以识别五种手势，这些手势可以用一只手完成，在按钮和口袋部位的准确率分别为89.2%和92.6%。
In this study, we show a new gesture recognition method for clothing-based gesture input methods using active and passive acoustic sensing. Our system consists of a piezoelectric speaker and a microphone. The speaker transmits ultrasonic swept sine signals, and the microphone simultaneously records the ultrasonic signals that propagate through the garment and the rubbing sounds generated by the gestures on the garment. Our method recognizes a variety of gestures, such as pinch, twist, touch, and swipe, by incorporating active and passive acoustic sensing. An important feature of our method is that it does not require a dedicated garment or embroidery embedded since our system only requires a pair of piezoelectric elements to be attached to the usual garment with a magnet. We performed recognition experiments of 11 gestures on the forearm with four types of garments made from different materials and recognition experiments of five one-handed gestures on the button of a shirt and the pocket of pants. The results of a per-user classifier confirmed that the f-scores were 83.9% and 95.9% for 11 gestures with four different types of garments and 5 gestures that were selected assuming actual use, respectively. In addition, we confirmed that the system recognizes five gestures, which can be performed with one hand, with 89.2% and 92.6% accuracy in the button and pocket sites, respectively.