TDOcc: Exploit machine learning and big data in multi-view 3D occupancy prediction

Chun Shan, Jian Zeng, Hongming Liu, Chuixing Chen, Xiaojiang Du, Mohsen Guizani

Abstract: Highlights•This research manuscript is an application and innovation of 3D occupancy prediction in the field of big data and machine learning.•In this work, we propose a novel framework called TDOcc, which applies big data and machine learning techniques to 3D occupancy prediction using multi-camera images. Compared to OccFormer, our approach offers two key advantages: (1) the use of dense occupancy labels enables effective dense scene occupancy inference and complete object estimation, and (2) the incorporation of historical feature information allows for effective alignment of temporal cues.•Furthermore, to alleviate the ill-posed nature of camera-based 3D occupancy and enhance perception performance, we introduce a voxel space enhancement module. Our approach demonstrates strong performance on the challenging nuScenes dataset, excelling in 3D occupancy prediction tasks.