基於電腦視覺方法利用無線網路訊號實現動作辨識

Abstract

由於無線網路訊號的普及性，低成本以及不侵犯隱私性，利用無線網路訊號實現動作辨識在近幾年來受到相當程度的重視。通道狀態訊息(Channel State Information, 以下簡稱CSI)是一種可從接收到的無線網路訊號預估出來的資訊，提供了詳細的環境如何影響訊號的訊息。我們觀察到CSI畫成圖片後，會由於不同動作而產生不同的材質變化，於是我們將不同動作預估出來的CSI都轉成類似頻譜的圖片，並從其抽取特徵值，接著訓練支持向量機進行分類，以實現動作辨識。實驗結果顯示，將CSI轉成圖片看待，在辨識7種不同動作能夠達到85%以上的準確率。 然而，實驗結果也顯示，預估出來的CSI存在著地點資訊，若無線網路訊號是從不同地點被錄下的，那麼原本訓練好的分類器將無法辨識該動作。於是，我們提出了一個基於奇異值分解(Singular Value Decomposition)的方法，能夠將地點資訊從CSI中移除，使得留下的資訊就是人的動作的影響。實驗結果顯示，在跨房間(地點)的動作辨識上，我們的方法在分類6個不同動作能夠達到90%的準確率。 我們的貢獻包括： 就我們所知，我們是第一個探討將CSI視為圖片，並且提出使用電腦視覺方法實現動作辨識的論文。 我們將不同對的天線視為不同信號通道，並且探討了不同特徵融合的方法，也達到了同一地點上不錯的動作辨識率。 我們提出的地點資訊移除方法，成功實現了跨房間的動作辨識。

Action recognition via WiFi has caught intense attention recently because of its ubiquity, low cost, and privacy-preserving. Observing Channel State Information (CSI, a fine-grained information computed from the received WiFi signal) resemblance to texture, we transform the received CSI into images, extract features with vision-based methods and train SVM classifiers for action recognition. Our experiments show that regarding CSI as images achieves an accuracy above 85% classifying 7 actions. However, from the experimental results, the CSI is usually location dependent, which affects the recognition performance if signals are recorded in different places. In this work, we propose a location-dependency removal method based on Singular Value Decomposition (SVD) to eliminate the background CSI and effectively extract the channel information of signals reflected by human bodies. Experimental results show that our method considering the correlation of CSI streams could achieve promising accuracy above 90% in identifying six actions even testing in five different rooms. Our contributions include: To our best knowledge, we are the first to investigate the feasibility of processing CSI by vision-based methods with extendable learning-based framework. We regard CSI of each Tx-Rx pair as a channel and investigate early and late fusion of multi-channels. Also, we achieve promising accuracy on action recognition on a specific location. We enable cross-room action recognition with the proposed location infomation removal method.