| Simultaneous Localization And Mapping(SLAM)technology provides a theoretical foundation for unmanned devices to operate in unknown environments.To promote the practical application of SLAM systems,this thesis focuses on the research of visual SLAM systems using commonly available monocular cameras.A monocular system can establish sparse constraint relationships and track the camera’s motion while inferring the environmental structure using multi-view geometry.However,this approach only yields a sparse point cloud,and it suffers from scale uncertainty.If deep learning techniques are employed to recover scale information from a single image,the resulting environment map can incorporate scale information.However,the structure of the point cloud can be severely distorted due to the blurring effect at the boundaries of the depth map.In addition,most traditional visual SLAM systems are based on static environments and cannot be well applied in dynamic environments.The constructed environmental point cloud does not have semantic information and cannot perform advanced semantic tasks.To address these issues,this thesis proposes a monocular visual SLAM system that integrates with deep learning.The main research content and innovation points of this thesis are as follows:1、In order to address the issue of lacking scale information in monocular systems,this thesis proposes a single-image depth estimation network based on unsupervised learning.Building upon existing monocular depth estimation networks,the network model is improved,and the overall optimization of the network is performed.The proposed network model achieves a reduction of over 50% in model size while maintaining accuracy.2、To address the issue of traditional visual SLAM systems being unsuitable for dynamic environments,this thesis proposes a semantic segmentation algorithm and integrates it into the visual SLAM system.The accuracy of the visual SLAM system in dynamic environments is improved by eliminating dynamic feature points.Compared to traditional visual SLAM systems on the TUM dataset,the proposed SLAM system reduces trajectory estimation error by approximately 18%.3、To address the problem of serious distortion of point cloud structures caused by the influence of blurred boundaries in depth maps when using deep learning to recover dense point clouds,this thesis proposes a point cloud fusion technique.This technique fuses sparse maps constructed by multi-view geometry and dense maps output by neural networks to obtain a dense point cloud map that is structurally robust and contains scale information.Semantic information is added to the map to construct a dense semantic map. |
