| Simultaneous Localization and Mapping(SLAM)for mobile robots is a mainstream method for robot environment perception and navigation.In the traditional visual SLAM methods,RGB-D information is of great importance.Visual SLAM systems can obtain data such as colors and depths from the RGB-D camera.Nevertheless,the process of obtaining depth data is always limited by factors such as distance and lighting.The monocular camera is of flexible structure and low price,and has high application value in practice.Accurately recovering three-dimensional information of the environment using a SLAM system and a monocular camera is a hot research direction.Besides,traditional SLAM method often uses the form of point cloud maps as reconstruction results,which is not sufficient for navigation.It is important to design a SLAM system that can reconstruct maps directly used in navigation practice.The actual robot work is often completed in complex dynamic scenarios,but the current SLAM method is often based on the assumption of static environment.Dynamic objects can disturb the process of feature point matching,so it is difficult to accurately reconstruct maps in dynamic environments.How to integrate semantic information into traditional SLAM methods to improve the accuracy of mapping in dynamic environments is also a hot issue at the moment.This thesis first designs and implements a new depth estimation algorithm.It solves the problem that a monocular camera cannot obtain the environmental depth information.This method is based on the encoder-decoder architecture,and modifies the structure of Deep Lab V3+.Firstly,we use high precision Res Net101 network combines a hole convolution model as the backbone network.Then,we use pixel-level classification results as the output to obtain an estimated depth map.In the process of depth estimation,we also design a dynamic frame skipping strategy,and optimize the loss function making the depth estimation result more smooth.Secondly,considering the navigation needs of the robot,this thesis designs and implements a complete monocular vision SLAM system based on the depth estimation method,which can reconstruct a three-dimensional grid map directly used for navigation.The design of the entire system is based on the visual inertial fusion SLAM method VinsMono.We obtain depth information using the monocular depth estimation method and recover a 3D point cloud map.This point cloud map helps to construct a 3D grid map that can accurately identify obstacle information.Finally,to improve the accuracy of mapping in dynamic environments,we summarize the existing methods,and propose a new scheme for reconstructing semantic maps in dynamic environments.The mapping scheme is based on the semantic segmentation method to semantically mark all objects in the scene.Based on the monocular vision SLAM system implemented in our task,the two-dimensional semantic information is associated with the three-dimensional point cloud,and the dynamic feature points are removed.At the same time,we maintain point cloud maps with semantics.When constructing the grid map,the probability update method is used to build a complete static grid map. |
