Markerless Human Tracking for Industrial Environments

Abstract

This thesis presents a markerless multiple-camera vision-based 30 human tracking method for industrial environments. The method can track humans in the vicinity of moving robots without using skin color cues or articulated human models. It is robust to self-occlusions and to partial occlusions caused by the robot. Foreground pixels corresponding to humans are found by background subtraction. A convex polyhedron enclosing the human(s) is generated online by bounding the foreground pixels in 30 space. Experimental results are included for a single person and multiple persons walking near a moving PUMA robot in a cluttered environment. Reliable tracking at 11.2Hz is demonstrated using four cameras and a Pentium 4 PC. The tracking data may be used for online robot collision avoidance.