Visual simultaneous localization and mapping using stereo vision with human body elimination for service robotics

Generally speaking, vision based simultaneous localization and mapping (V-SLAM) systems have to avoid obstacles especially for the people moving around in the environment because it will cause temporary landmarks captured by camera, and thus will increase computational loading and estimative error in the V-SLAM system. However, in some situations, we cannot always evacuate people when the robot is collecting data. In this paper, we propose a V-SLAM system which consists of human body elimination and verifying that the human body filter can decrease estimative error to the system. We employ FAST machine learning approach to perform corner detection which is faster than most of the other feature detection methods with similar quality for detecting the location of landmarks. We al so implement “landmark buffer” to find robust landmarks in a frame. We also detect where the human body appears in a frame and eliminate the landmark features extracted from human body. Through landmark buffer and human body elimination, our approach derives accuracy estimation of the robot pose. This proposed approach has been successfully demonstrated through experimental verification and the results are summarized in the conclusion.