@techreport{oai:ipsj.ixsq.nii.ac.jp:00218009,
 author = {堤, 日向 and 武中, 紘輝 and 小林, 慧 and 近藤, 圭 and 長谷川, 達人 and Hyuga, Tsutsumi and Koki, Takenaka and Satoshi, Kobayashi and Kei, Kondo and Tatsuhito, Hasegawa},
 issue = {22},
 month = {May},
 note = {センサベースの行動認識では深層学習手法が多く利用され,認識精度向上に貢献している.モデルの入力には主に加速度やジャイロセンサを用いるが,加速度センサデータを周波数スペクトルに変換して用いることもある.しかし,周波数特性に着目したデータ拡張はこれまで深く議論されていない.本研究では行動認識における各行動を推定する際に重要な周波数を強調するフィルタとアンサンブル学習を用いた行動認識手法を提案する.提案手法の実現に向け,加速度センサデータに対し一部の周波数帯をマスクし,そのデータを用いて精度を比較することで各行動の重要な周波数を実験的に明らかにした.提案手法の有効性を示すために,訓練時の強調フィルタの有無,テスト時の強調フィルタの有無,アンサンブルの有無を組み合わせて精度を比較した.その結果,訓練時とテスト時に周波数帯強調フィルタを適用し予測結果をアンサンブルすることで認識精度が最も高くなり,提案手法の有効性を示した., Deep learning methods are widely used in sensor-based activity recognition, contributing to improved recognition accuracy. Acceleration and gyro sensors are mainly used as input to the model, and sometimes accelerometer data is converted to a frequency spectrum. However, data augmentation focusing on frequency characteristics has not been deeply discussed. This study proposes an activity recognition method that uses an ensemble learning and filters that emphasize the frequencies important for estimating each activity. In order to realize the proposed method, we experimentally revealed the important frequencies of each behavior by masking some frequency bands in the accelerometer data and comparing the accuracy using the masked data. To demonstrate the effectiveness of the proposed method, we compared the accuracy of the method with and without enhancement filter during training, with and without enhancement filter during testing, and with and without ensemble learning. The results showed that applying the frequency band enhancement filter during training and testing and ensemble achieved the highest recognition accuracy, indicating the effectiveness of the proposed method.},
 title = {周波数帯強調フィルタとモデルアンサンブルを用いたセンサベースの行動認識},
 year = {2022}
}