![]() | Only 14 pages are availabe for public view |
Abstract Effective perception of the environment is crucial for autonomous driving, necessitating a perception system capable of capturing detailed 3D information about surrounding objects, including their dimensions, locations, and spatial orientation. In the realm of perception systems, deep learning has gained widespread adoption, leveraging image features from cameras to derive semantic information. This thesis introduces the MonoGhost network, a lightweight Monocular GhostNet deep learning approach designed for comprehensive 3D object property estimation from a single monocular image frame. Distinguishing itself from other methods, the proposed MonoGhost network prioritizes the initial estimation of reliable 3D object properties through an efficient feature extractor. This encompasses the orientation and dimensions of the 3D object, exhibiting notably minimal errors in dimension estimations compared to alternative networks. Leveraging these estimations alongside translation projection constraints derived from 2D detection coordinates enables the prediction of a robust and dependable Bird’s Eye View bounding box. Experimental results affirm the superior performance of the proposed MonoGhost network, outperforming state-of-the-art networks in the Bird’s Eye View benchmark of the KITTI dataset with scores of 16.73% on the moderate class and 15.01% on the hard class, all while maintaining real-time processing requirements. |