System for detection and classification of obstacles through data-fusion of lidar and RGB cameras (ASGARD-CLASS)

The ASGARD-CLASS software system is designed for obstacle detection in lidar data in a point-cloud format, followed by classification of the obstacle type into several classes. The system will find use in mobile robotics, where the data created by this program can serve as a simplified and easily searchable map of the robot's surroundings. It enables obstacle detection, floor filtering and filtering of solid parts of the environment (e.g. walls) using advanced AI-based segmentation techniques. The classification of the obstacles detected in this way is then performed on a data fusion of lidar data and data from RGB cameras, which are annotated using a neural network based on a pre-trained YOLO network, which was further trained on its own dataset from an industrial environment. The system is compatible with the ASGARD-NAV navigation, planning and mapping tool, which can use information about the obstacle class for better trajectory planning.