DeepFusion encoder for unsupervised monocular metric depth estimation

Monocular depth estimation is a fundamental task in computer vision, with significant applications in autonomous driving and robotics. However, accurately estimating depth from a single image remains challenging due to the absence of direct depth cues. In this article, we present a novel self-supervised monocular depth estimation framework that integrates neural networks with multiscale 3D spatial loss to directly predict metric depth, thereby addressing the scale ambiguity inherent in monocular setups. Our approach introduces three key innovations: (1) an innovative integration of attention mechanisms within the hierarchical context encoding structure, which facilitates adaptive depth estimation; (2) the incorporation of kinematic scaling loss into continuous time frames for gradient backpropagation, significantly alleviating the scale ambiguity issue associated with monocular depth estimation; and (3) the construction of the University of Malaya Autonomous Driving (UMAD) dataset—an entirely original urban autonomous driving dataset created using Frequency-Modulated Continuous-Wave Light Detection and Ranging (FMCW LiDAR)—alongside methods for training, testing, and evaluation. Extensive experiments conducted on both the Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago (KITTI) and UMAD datasets demonstrate that our method achieves substantial improvements in accuracy as well as real-time performance for depth predictions. Practical applications in monocular 3D perception further validate the effectiveness and robustness of our approach within autonomous driving scenarios.