MoveEnet: Online High-Frequency Human Pose Estimation with an Event Camera
Gaurvi Goyal, Franco Di Pietro, Nicolò Carissimi, Arren Glover, Chiara Bartolozzi
Abstract
Human Pose Estimation (HPE) is crucial as a building block for tasks that are based on the accurate understanding of human position, pose and movements. Therefore, accuracy and efficiency in this block echo throughout a system, making it important to find efficient methods, that run at fast rates for online applications. The state of the art for mainstream sensors has made considerable advances, but event camera based HPE is still in its infancy. Event cameras boast high rates of data capture in a compact data structure, with advantages like high dynamic range and low power consumption. In this work, we present a system for a high frequency estimation of 2D, single-person Human Pose with event cameras. We provide an online system, that can be paired directly with an event camera to obtain high accuracy in real time. For quantitative results, we present our results on two large scale datasets, DHP19 and event-Human 3.6m. The system is robust to variance in the resolution of the camera and can run at up to 100Hz and an accuracy 89%.