3D terrain reconstruction of construction sites using a stereo camera

Abstract

Abstract This paper presents a fast and robust three-dimensional (3D) terrain reconstruction system that uses a stereo camera. Local feature-based dense 3D reconstruction consists of two major steps: matching correspondence points and dense 3D reconstruction. In the matching step, the descriptor is an important component, as its properties significantly affect the precision of the matching. Furthermore, matching is the most time-consuming step. In this paper, correspondence points are found using multi-scale descriptors (MSDs) because of their robustness and computational efficiency. A two-stage cascade matching method suitable for MSDs is also proposed. In the dense 3D reconstruction step, a probabilistic model is proposed for dense reconstruction that provides high precision through the use of robustly matched correspondence points and computational efficiency by narrowing the search range using coarsely inferred disparity values from precisely calculated triangle meshes. To collect experimental data, a prototype stereo camera system is also built that is mounted on the front of an excavator. This paper concludes by comparing the proposed dense 3D reconstruction with different types of dense 3D reconstruction methods in terms of processing time and similarity to the shape of the terrain. The results from an evaluative experiment show that MSD-based dense 3D reconstruction is suitable for various autonomous control applications in construction sites where computation time and precision are vital.