Estimation of range-gated depth in autonomous driving with RGB geometric priors

Abstract

Existing deep learning-based 3D range-gated imaging methods, like Gated2Depth, still lack the ability to fully exploit color cues, have high sensitivity to specular noise, and produces undesirable local artifacts in autonomous driving situations. To overcome these limitations, we suggest RGD-Net, a new depth estimation network that incorporates RGB geometric priors. The network employs three consecutive range-gated frames as its input. Synchronized RGB images are projected and aligned with the gated coordinate system using camera calibration parameters to provide geometric priors such as scene edges, texture gradients, and spatial structures. Based on the encoder-Transformer-decoder architecture, this framework employs a cross-modal fusion strategy with a cross-attention mechanism to effectively combine depth features from gated images with edge and texture information from RGB images. Using the global modeling ability of Transformer, our model can capture long-range spatial dependencies and reduce local noise interference. Edge alignment and relative depth consistency constraints are further designed to facilitate the learning of depth boundaries and scene geometric structures. Experimental results on the Gated2Depth dataset demonstrate that compared with the baseline, the proposed method reduces MAE by 42.84% and RMSE by 29.00% in daytime scenarios, while boosting δ1 accuracy by 18.89 percentage points. The generated dense depth maps exhibit clearer boundaries, effectively correcting erroneous predictions in depth-discontinuous regions and improving the integrity and accuracy of object contours. Experimental results indicate that RGD-Net is promising for range-gated 3D imaging in autonomous perception tasks.