Predicting Three-Dimensional Gait Parameters with a Single Camera Video Sequence

Abstract

Human gait reflects biomedical conditions and thus can potentially be used for identification. With the increasing utility of CCTVs for surveillance, there have been various attempts to recognize persons using gait image sequences from a single camera. We investigated the accuracy of estimating body segment lengths and joint angles during gait calculated from a video sequence using a gait database. We recruited 30 subjects and collected motion capture data during walking and extracted the trajectories of 17 body points. Principal component analysis (PCA) was applied to the collected gait. We implemented full gait cycle-based (FGC) PCA and gait-phase-specific (GPS) PCA. Three-dimensional poses were estimated from gait event frames using FGC-PCA and GPS-PCA. The estimated poses in discrete gait event frames were interpolated to estimate motion during a full gait cycle. The body pose from GPSPCA was less sensitive to camera angles and smaller errors compared to FGC-PCA. The segment lengths of the upper arm (r=0.79), lower arm (r=0.63), upper leg (r=0.86), and lower leg (r=0.81) were highly correlated with the lengths obtained from the motion capture data. Three-dimensionally reconstructed human motion can reveal personal biometric information and has the potential to be used for human identification.