RDDepth: A Lightweight Algorithm for Monocular Depth Estimation

Current monocular depth estimation algorithms have complex structures and high computational costs, making it difficult to deploy on embedded devices. To address these issues, this paper proposes a lightweight monocular depth estimation algorithm: RDDepth, which aims to achieve fast and accurate depth estimation on embedded devices. Firstly, RegNet is adopted as the Backbone to balance the accuracy and computational cost of the algorithm; Secondly, DenseASPP is used to replace ASPP in the Neck to enhance the accuracy of depth estimation; Thirdly, channel pruning is applied to the Neck and Head of the algorithm to reduce the parameters and computational complexity; Finally, to improve the accuracy of depth estimation, an improved loss function is proposed. Experiments are conducted on the public dataset KITTI, and the algorithm is deployed on the NVIDIA Jetson Xavier NX. Experimental results show that, compared to the baseline, RDDepth increases the FPS with FP16 accuracy by 4.34 times to 44.01 FPS, reduces the parameters by 4.7 times, and lowers the GFLOPs by 8.17 times, while also reducing the root mean square error (RMSE) by 0.34. In all, RDDepth achieves significant improvements in all aspects and can perform realtime and accurate monocular depth estimation on embedded devices.