Gait Recognition from Drone Videos

Abstract

Gait recognition is important in humanitarian assistance and disaster relief (HADR) missions. It is challenging to detect human pose and recognize the gait from a drone video in distance due to low-resolution in human-figures, dynamic pose transition, and figure-to-figure and figure-to-background interaction. Conventional pose estimation methods often fail to register proper gaits from drone footages. In this study, In this study, we explore the Adaptive Pose Estimation (APE) method that incorporates drone flight data and camera pose estimation to improve the pose estimation in drone videos. We also explore the novel methods to estimate the drone camera pose with Google Earth 3D model and to reduce the “phantom poses” with “amputation filter.” Our results show that the pose estimation accuracy increases 36% and the recognition rate for six gaits reaches 91.7% on average.