Loop-Closure Detection In Visual SLAM And Path Planning In Mobile Robots

The rapid development of science and technology promotes a series of high-tech products such as AR/VR,robots,and unmanned vehicles etc.Such high-tech products require the support of many technologies.This thesis focuses on two issues: loop-closure detection in the Visual Simultaneous Localization and Mapping(V-SLAM)system and Path Planning.To cope with the drifting problem of robot pose estimation over time,loop-closure detection technique is proposed to achieve globally consistent trajectories and maps.In the process of experimenting and analyzing the bags-of-visual-word(BoVW)model,it is found that the model based on features performs well only in certain scenes.At the same time,the process of extracting features is time consuming.Inspired by the idea of the BoVW model,the proposed method employs the convolutional neural network(CNN)to extract the features of images.The features is more universal than those extracted by the BoVW method.The experimental results demonstrate that the outputs from the POOL5 layer are the best feature descriptors.Then the Euclidean distance between the image features is calculated to obtain a similarity matrix.In order to deal with the active mapping problem of robot,boundary detection algorithm based on local and global Rapidly-exploration Random Tree(RRT)is proposed to detect the boundary points between the known region and the unknown region quickly and efficiently.After obtaining the map,an algorithm that combines biased RRT and DWA is proposed to obtain the velocity command and global path.The proposed method is verified by the tools embedded in Robot Operating System(ROS).Compared with A* and Dijkstra,the path planning algorithm in this paper is much smoother and faster.