@article{oai:ipsj.ixsq.nii.ac.jp:00175062,
 author = {宮田, 章裕 and 伊勢崎, 隆司 and 中野, 将尚 and 石原, 達也 and 有賀, 玲子 and 望月, 崇由 and 渡部, 智樹 and 水野, 理 and Akihiro, Miyata and Takashi, Isezaki and Masanao, Nakano and Tatsuya, Ishihara and Reiko, Aruga and Takayoshi, Mochizuki and Tomoki, Watanabe and Osamu, Mizuno},
 issue = {10},
 journal = {情報処理学会論文誌},
 month = {Oct},
 note = {車椅子ユーザの移動を妨げる“バリア”に関する情報を収集することが，社会的に求められている．しかし，既存手法では，電柱や道路工事などの路上物バリアを発見することは困難である．我々は，路上物バリアをよける・引き返すといった動作を構成する車椅子操作を推定できれば，操作の推定結果をもとに路上バリアの存在を検出できると考えている．このとき，操作推定は，多くのユーザが持つスマートフォン内蔵の加速度・角速度センサを利用して行えることが望ましい．ところが，車椅子操作時に発生する加速度・角速度は，同じ操作であってもユーザ・操作習熟度・状況によって大きく振舞いが異なり，単一の推定器では精度良く推定できないという問題がある．そこで，我々は，ユーザの直近の移動能力を推定した後に，その移動能力に適した推定器で操作内容を推定するアプローチを提案する．ディープラーニングを用いて提案コンセプトを実装し，ユーザの実走行データに基づいて構築した約25万件のデータセットを用いて実施した検証実験では，提案手法が前後進・右左折・停止などの操作を高精度に推定できるだけでなく，段差・傾斜の検出にも有効であり，適合率・再現率はともに98.9%とベースライン手法を大きく上回った．提案手法は，バリア情報の収集に大きく貢献しうるものと考えられる．, It is socially desired to collect information on “barriers” that prevent free movement of wheelchair users. But, existing techniques have difficulty in detecting barriers on the road such as power poles and roadworks. We consider it is possible to detect barries on the road by estimating wheelchair motions that constitute actions such as changing the course or turning back due to barries. Here, it is deemed disirable to estimate wheelchair motions using the accelerometer/gyroscope in the smartphone that most users already have. However, a single estimator fails to estimate wheelchair motions with a high degree of accuracy. Because behaviors of acceleration and angular velocity vary according to the user, ability of controlling the wheelchair and situation even if the motion to be estimated is same. To address this probrem, we propose a wheelchair motion estimation method that consists of two steps; in the first step, esitimates the user's ability of controlling the wheelchair, and in the second step, estimates the wheelchair motion using the suitable esitimator for the ability. To conduct an evaluation task, we made a prototype system based on Deep Learning, and constructed a dataset of two hundred and fifty thousand motions extracted from actual wheelchair driving data. The result shows that the proposed method can estimate not only motions such as moving forward/backward, turning right/left and stop, but road conditions such as a step and slope, with an accurascy of 98.9 percent that greatly exceeds that of the baseline method. The proposal method is considered to contribute collecting barrier information.},
 pages = {2316--2326},
 title = {直近移動能力を考慮した車椅子操作推定モデル},
 volume = {57},
 year = {2016}
}