深度情報のない歩行画像からのCNNによる床反力推定

―汎化性能の検討―

抄録

Although floor reaction force data is useful for gait analysis, it can only be measured in specialized facilities because it is often measured using installed force plates. This report examines a simple and highly accurate three-directional floor reaction force estimation method based on images taken with devices that most people have. By using only captured images for estimation, we believe that floor reaction force estimation can be easily performed without the need for specialized knowledge, sensors devices, skeletal information including uncertainties obtained from images, and statistics for each body segment. CNN, a type of deep learning, is used for this learning. Generalization performance for untrained subjects is evaluated by cross-validation using data from five subjects. The estimation results were accurate with errors of 8% of body mass in the vertical direction, 2% of that in the medial-lateral direction, and 3% of that in the anteriorposterior direction for a normal gait. Accordingly, the generalization performance of the proposed method was confirmed.