Object Recognition Findings in a Built Environment

Abstract

In order to prepare the point cloud for a feature-based process of planning and simulation, the objects it contains must first be recognized and a CAD-based structure must be created. The great challenge in every process chain is to implement the desired function with minimal effort and losses. As one of the alternatives, the object recognition is selected to implement the layout planning of a factory in 3D. In this chapter, it is demonstrated how to embed the object recognition in the process of virtual 3D layout planning in a built environment as well as which findings and results can be expected. The aim of this chapter is to investigate and evaluate the usefulness of a realistic 3D virtual factory model in factory layout planning primarily for Digital Twin based on object recognition. This is addressed by a practical study of how existing methods for data acquisition and processing can be concatenated and, subsequently, applied under real industrial constraints and conditions. During this study, realistic 3D layout models are created using point clouds acquired by commercial terrestrial laser scanner and prepared for object recognition with Convolutional Neural Networks considering the strict data quality requirements. The selection of models was discussed and the results were evaluated in industrial workshops with engineers involved in the layout planning and machine operators that will work within the production system. Seamless, robust, (semi-)automatic workflow of primarily standard, modular components with low user assistance was of particular interest. This chapter is concluded with the discussion of the achieved results, the solution alternatives, and the present approaches how to extend the utilization, improve, and simplify the entire process.