On-board 3-D SLAM for AGV localization

Abstract

Human detection is an essential task for a SLAM system in dynamic environment, such as warehouses, where the robot shares its workspace and interacts closely with operating personnel. Being aware of the real-time locations of humans in the scene is the basis for safe operation of the system. This final year project presents a 3D LiDAR SLAM system with an efficient human classification function that utilizes human geometry and state-of-the-art machine learning techniques to accurately identify human clusters in complex 3D point clouds. The performance of the system is evaluated with firsthand data from the Delta-NTU Corporate Laboratory for Cyber-Physical Systems. Experiments show the combination of the anthropometric and SVM classifiers produces decent human classification results in warehouse environment with medium to high complexity.