Research And Simulation Of RGB-D SLAM Gorithm Based On Self-Supervised Image Depth Estimation

In recent years,the demand for high-precision and robust ubiquitous positioning and navigation is increasing.Especially in urban vehicle positioning,patrol robot positioning and other fields,this demand has broad application prospects.Traditional solutions mainly rely on global navigation satellite system and inertial navigation system.The former affects the positioning accuracy under the occlusion of the city,while the latter is affected by hardware noise,and the system diverges with time without the assistance of other systems.With the rapid development of computer vision technology,monocular vision Slam(simultaneous localization and mapping)has become an important means to achieve accurate and autonomous positioning.However,the current monocular vision slam system still faces the challenges of scale uncertainty and insufficient robustness.In addition,monocular cameras cannot directly obtain scene depth information.Therefore,improving the robustness of visual SLAM algorithm and the accuracy of monocular depth estimation is of great significance for achieving accurate autonomous localization.In order to solve the above problems,this paper deeply analyzes the existing algorithm framework,and proposes a dense SLAM algorithm based on rgb-d image input,aiming to achieve more reliable positioning.For the limitation that monocular cameras cannot directly obtain scene depth information,this paper adopts a self supervised learning framework,which can train the depth prediction network relying on monocular video,so as to achieve high-precision depth estimation of monocular images.The monocular depth information obtained by prediction can effectively support the monocular system to achieve accurate autonomous positioning.The primary contributions of this study are as follows:(1)The proposal of a self-supervised monocular depth estimation network.To imbue the network with scale-awareness and robustness to normal lighting conditions,two sets of networks are utilized.One set is dedicated to estimating image depth information,while the other focuses on motion estimation.Through the construction of well-defined loss functions,these two network sets mutually learn and extract image features,thereby achieving more accurate,robust,and scale-stable selfsupervised image depth estimation tasks.Comparative experiments reveal that the algorithm effectively outlines the contours of critical objects and demonstrates a favorable portrayal of their depth relationships;(2)The establishment of a visual SLAM system based on RGB-D image input.Compared to monocular systems,RGB-D images provide supplementary depth information,enabling more precise localization.Concurrently,the utilization of RGB-D information facilitates dense mapping functionality.Comparative experiments demonstrate that the introduced SLAM system,equipped with additional depth information,achieves superior localization accuracy on public datasets.