Depth-prior-based range-gated 3D imaging method: integration of optical property into visual learning

Abstract

3D Range-gated Imaging (3DRGI) has great potential for long-range detection in adverse weather conditions. Recently, vision-guided 3DRGI has brought new perspectives to this area as it overcomes hardware limitations and greatly increases flexibility. However, existing vision-guided methods do not consider the optical properties of range-gated imaging, which results in low accuracy. This paper proposes a depth-prior-based 3DRGI method to combine the advantages of optical-based and vision-guided methods. In this method, depth-prior is firstly deduced from principles of range-gated imaging and provides effective depth signals. Then, adaptive depth intervals are estimated using statistical methods, and a depth-prior-guided loss function is designed. The integration of the depth-prior-guided loss function within a vision-guided model enables focused attention on pixels with depth estimations that are inconsistent with the depth prior, thereby refining the overall accuracy of the depth map. To prove the feasibility of the proposed method, comparison experiments and ablation studies have been performed. The results show that the root mean square error improved by more than 12% under adverse weather conditions.