비전 기반 3차원 자세 추정을 이용한 인간 로봇 충돌 방지 시스템

Abstract

As human-robot collaboration becomes increasingly prevalent in medical and industrial settings, the risk of collisions is rising due to shared workspaces. Traditional approaches rely on stopping mechanisms or external sensors, which may limit system adaptability. This study proposes a camera-based method utilizing a convolutional neural network (CNN) to estimate the 3D joint coordinates of robots and humans within a unified camera coordinate frame. It estimates the 3D positions of the robot's seven joints with an average error below 5.5 mm, enabling real-time collision risk assessment through joint distance computation. This vision-based approach provides precise spatial awareness without requiring additional external sensors, enhancing safety and operational efficiency.