Design of a Resource-Oriented Framework for Point Cloud Semantic Annotation with Deep Learning

摘要

With the development of mobile mapping systems, point cloud is becoming an important data type to build the three-dimensional geospatial information. In particular, some recent applications such as robot navigation, self-driving drones/vehicles, and construction robotics need three-dimensional high-definition maps within cm-level accuracy to make a decision of action by themselves. These maps are based on point cloud data from three-dimensional surveys of geographic features on the Earth by laser scanning and photogrammetry. Point clouds provide a flexible and accurate geometric information for reconstructing 3D shapes of not only indoor but also outdoor features. However, it expends time on creating a map due to many manual operations of processing. In this paper, we design an interoperable framework of semantic annotation for point cloud data. Semantic annotation of point cloud is the process of attaching conceptual label information such as cars, buildings, desks, or people to each point in a set of data. This process takes first priority to create training data set for machine learning and deep learning tasks. An efficient annotation tool for 3D point cloud data is crucial to improve the quality of deep learning models and getting better results in deep learning would be helpful to reduce the manual processing. Even though there are already a number of tools, they are usually standalone and tightly coupled systems. This is likely to lead to a problem of limited interoperability of data and process between systems. This paper introduces a resource-oriented design for the annotation framework of point cloud data on the Web with the ability of quick adaptation of various machine learning techniques.