Accurate Monocular Road Depth Estimation for Ground Vehicles Using Suspension Feedback

Road depth estimation plays a crucial role in vehicle chassis control and autonomous driving. Monocular depth estimation, due to its low cost, energy efficiency, and ease of deployment, remains an important area for research. However, challenges persist in improving the accuracy of monocular depth estimation, reducing the impact of errors, and ensuring the convergence of errors over long-term operation. This paper proposes an improved monocular depth estimation approach that integrates deep learning techniques with feedback from vehicle suspension states. Under the DispNet architecture, we replace traditional neural networks with the Standard Nonlinear Operator Form (SNOF) and incorporate vehicle suspension information to build a delayed state observer for real-time depth estimation error compensation. State compensation is achieved through solving Linear Matrix Inequalities (LMI), ensuring error convergence during extended operation. Extensive real-world experiments using signal-collecting vehicles demonstrate that the proposed method exhibits excellent generalization capabilities across diverse environments and lighting conditions.