Bilateral Thigh Data Fusion in Convolutional Neural Networks and the Optimal Input Set for Gait Phase Prediction of Walking Along Curved Paths

Although walking along curved paths is common in daily life, most gait studies have focused on walking along straight paths. Thus, this study thoroughly investigated the use of a convolutional neural network (CNN) for gait phase (GP) estimation during curved walking. First, the thigh angles were acquired in three-dimensional space. Subsequently, several combinations of the angles were used as the input sets of the CNN. Second, three CNN models (i.e., no-fusion, late fusion, and early fusion) were created, differing in how left and right thigh angles were integrated. When sufficient computational resources were available (e.g., workstation), the early-fusion model using all angular components achieved the highest prediction accuracy. In contrast, under computational constraints (e.g., microcontroller units), the early-fusion model with inputs limited to sagittal and transverse angles provided the best trade-off between accuracy and inference speed. Moreover, this study revealed that GP prediction accuracy strongly depends on the curvature of the path. The results of this study can be used to design a CNN model for walking along curved paths for various applications such as clinical studies, rehabilitation, and wearable robots.