Methodology [原文图片,侵删]
System Design 上的改进
1)文章利用OpenWrt, 一款 lightweight Linux OS 作为IoT platform 的OS来开发平台,用于收集WiFi CSI的数据, 相较于传统的Intel 5300 NIC tool,该平台将subcarriers数目从30拓展到了114 (更多的subcarriers意味着更多的CSI info)
2)文章利用CSI phase data来做gestures learning, 因为认为phase比amplitude更sensitive
Siamese Recurrent Convolutional Network
Model
Inputs: two gesture stream samples
Loss function:
1) Pairwise loss: The pairwise loss guarantees that the distance between the same samples
is restricted within m − b while the margin of samples from different classes exceeds m + b.
2) MMD: 为了使得model具有更好的transferability, 可以有效处理来自不同domain的samples,引入multiple kernel variant of maximum mean discrepancies (MK-MMD)。MK-MMD bears the unique property for measuring two distributions of samples from different domains
Total loss:
Algorithm
利用one-shot learning 做gesture recognition
当model已经利用足够的samples训练之后,我们可以在新的环境中为每组gesture(不考虑新的gestures)收集一个或者几个samples并进行label,之后可以通过比较test sample和each class的Euclidean distance进行分类
Experiments & Evaluations
Cons:
1) 考虑的gestures classes 较少 (6组?)
2) 只考虑了新的环境下同样classes的testing,未来可以多考虑一些新的不同的gestures