Depth prediction method based on monocular images and sparse point cloud fusion

Monocular depth estimation suffers from inherent limitations in directly acquiring depth information, resulting in constrained prediction accuracy. This paper proposes a depth prediction method based on the fusion of monocular images and sparse point clouds, combining joint calibration with a multimodal fusion algorithm. The joint calibration method employed in this study utilizes a graph neural network to achieve the registration of image and point cloud key points through local feature matching, thereby reducing depth prediction errors at the dataset level. After registration, a residual neural network-based multimodal fusion algorithm is adopted to extract and fuse features from the monocular image and sparse point cloud, ultimately outputting a depth image. Experimental results demonstrate that, compared with monocular image depth estimation, the proposed method improves depth prediction accuracy from 57% to 93% on the KITTI dataset and from 55% to 86% on a self-constructed tank target dataset. This approach achieves high-precision depth prediction for outdoor environmental targets, providing a reliable solution for enhancing depth estimation accuracy.