WiFi is a powerful sensing medium Gait Identification Gesture Recognition
Limitations Require collecting training data in all sites Only support for a handful of users 同一个人、同一手势在不同地方的特征表现非常不同（原因：multipath effect）。图为同一个人的手势在两个不同地点的特征。收集数据需要耗费大量时间 User数目增多时，准确率下降明显。
Goals Scaling to new environments Consistent performance for larger problem sizes 仅在一个地点收集数据，但是通过一些方法得到多个环境下的模型，达到跨站点的效果 分类的类别数增加依然能保持很好的效果（步态识别中的用户数目，手势识别中的手势数目）
y = f(x) Roaming WiFi metrics Input signal metrics Learning how does the environment affect the WiFi signal Input signal metrics for site A Translated signal metrics for site B y = f(x) 不同地点的数据扩展 Roaming model
Using multiple sensing models Dynamically choose which sensing model (expert) to use Sensing model to use Input WiFi signal y = g(x) 问题规模增大时的解决方案：采用expert selector选择不同的模型 Expert selector
How do the two parts work together? Roaming model WiFi training data for new sites 1. 训练一个roaming model，offline WiFi data
Model Input Noise Removal Feature Selection WiFi data
Artificial neural network based data roaming Roaming model WiFi training data for new sites WiFi data
Roaming model 7-layer MLP Training data Transfer learning data collection site deployment site Data collection time cost ~14 hrs Transfer learning Shared params (first few layers) between ModelA_B and ModelA_C Reducing data collection cost largely
Classification based model selection Roaming model WiFi training data for new sites WiFi data
Sensing model chosen based on input signal characteristics kNN classifier Sensing model i Translated WiFi signals … Sensing model N
Prior art Gait Identification Gesture Recognition WiWHO WiFiU WiG WiAG Tens of participants Handful of gestures
Evaluation Setup Environments: 3x sites: low, medium and high multipath Activities: Gait identification (100 users) and gesture recognition (40 gestures) Wireless channel metrics: CSI and RSSI
Consistently good performance
Roaming WiFi data improves the performance
Conclusion Problem: Wifi sensing is great, but limited to small-scale deployments Solutions: Machine learning for roaming training data and scalable performance Results: 4x improvements on accuracy much larger problem sizes
Thank you