920MHz帯電波受信レベルのウェーブレット解析に基づいた機械学習による屋内在室不在判定手法

防犯対策や高齢者の見守り，働き方改革による労働者の勤務状況の把握などの必要性が高まっている．これらの対策としてUHF帯テレビ放送波や920MHz帯電波など，様々な周波数帯を用いたヒト検知手法を提案している．920MHz帯はスマートメータやセンサネットワーク，HEMS（Home Energy Management System）などのIoT（Internet of Things）で用いられるため，住宅で複数機器がネットワークにつながって利用されることが考えられる．そこで本研究では，複数の920MHz帯無線端末のRSSI（Recieved Signal Strength Indicator）によるヒトの在室不在判定を実現することを目的として，送信機1台，受信機3台を用いた測定システムを構築して，屋内環境においてRSSIを測定した．測定したRSSIにウェーブレット変換を適用し，得られたウェーブレット係数を用いて在室不在判定を行った．判定手法として，SVM（Support Vector Machine），k近傍法，閾値決定法を用いた．ウェーブレット変換の周波数に対する判定精度の評価を行い，各判定手法に適したウェーブレット変換の周波数の選定を行った．また，選定した周波数のウェーブレット係数を用いて学習器への入力形式を複数検証し，在室不在判定性能向上を実現した．

There is a growing need for security measures, watching over elderly people, and monitoring the work status of workers through work style reforms. We has proposed a human detection method using radio waves in various frequency bands, such as UHF band broadcast waves and 920MHz band radio waves. It is expected that multiple wireless devices will be connected to the network in the future，because the 920MHz band is used for IoT (Internet of Things) such as smart meters, sensor networks, and HEMS (Home Energy Management System). In this study, we aimed to detect presence/absence of a human in a room, and developed the human detection system using RSSI (Received Signal Strength Indicator) of multiple wireless modules. A measurement system was constructed with three receivers and one transmitter in the 920MHz band. A wavelet transform was applied to the measured RSSI, and the resulting wavelet coefficients were used to detect the presence/absence of a person in the room. SVM (Support Vector Machine), k-nearest neighbor method, and threshold determination method were used as detection methods. The detection performances were evaluated by changing the frequency of the wavelet transform, and the frequency of the wavelet transform suitable for each detection method was selected. By use of the wavelet coefficients of the selected frequencies, we verified the performance of the human presence/absence detection in the room with several input to the learning machine, and found that the learning machine based detection method improved the performance.