A Real-Time and Lightweight Monocular 3-D Object Detector on CPU-Based Edge Devices for UGV's Indoor SLAM Systems

Abstract

In this article, we introduce RLMono3d, a real-time monocular 3-D object detector designed to run efficiently on edge devices with just one CPU. This approach is specifically applied to the indoor simultaneous localization and mapping (SLAM) system on the unmanned ground vehicle (UGV) platform. The spatial projection model for 3-D cuboids is proposed using projective geometry to determine the topological relationship between the cuboid projection and the bounding box (Bbox). Utilizing the theory above, we propose a real-time lightweight monocular 3-D object detector that is designed to be efficient and lightweight by using 2-D object and line segment detectors to generate 3-D object Bboxes. Furthermore, this approach is integrated into the SLAM system to create a complete semantic SLAM system, which is a typical application of our approach. Experimental results on multiple indoor benchmarking datasets show that the proposed monocular 3-D object detection algorithm outperforms benchmark methods. To validate our approach on the UGVs, we also conducted experiments using the quadruped robot platform. The results demonstrate that our approach can operate on edge devices of UGV platforms and accurately detect 3-D objects in real time.