A Novel Deep Learning Based Method for Automated Foot Motion Measurement During Walking

Abstract

Accurately tracking of multi-segmental foot motion during gait is critical for clinical diagnosis and rehabilitation interventions related to the foot and ankle disorders. Recently, a Point-cloud-based Foot Analysis (PFA) method has been developed for rapid acquisition of foot motion data in the form of point clouds, but the manual labeling of different foot segments prone to errors and limit the efficiency in subsequent joint kinematic analysis. Therefore, this study focused on a deep-learning-based modeling method for automated foot segmentation and kinematic analysis. In our experiment, we trained PointNet++ using 109 manually annotated foot point cloud data, and applied the resulting best model to the testing set. The results demonstrate that the model was able to accurately divide the foot point-cloud data into the five segments, where the Overall Accuracy (OA) and mean Intersection over Union (mIoU) of the testing set are 92.1% and 85.4%, respectively. Although unsmooth boundaries were still observed in each segment, the obtained kinematic data exhibits good consistency with those of the manual segmentation. To conclude, the novel automated foot motion measurement method based on the PointNet++ segmentation model offers higher efficiency as a more promising tool and contributes to the establishment of the large-scale foot-ankle kinematic database towards clinical applications.