Robust Visual Navigation And Environment Exploration With Mobile Robot

In the past ten years,AI has developed at an unprecedented speed and has shown amazing application potential in various areas.In the field of mobile robots,smart mobile devices are required to have the ability to perceive the external environment,localization in real time and navigation in the modeled environment.Among them,visual environment modeling and visual localization based on image and video sequences have attracted a lot of attention in recent years because of its low cost,rich information and low requirement of infrastructure modification.With the development of many scholars in this field,some problems of visual navigation have accumulated considerable theoretical results.Especially in the VSLAM area(Visual Simultaneous Localization and Mapping),at present,in the accuracy of mapping and localization,a series of landmark achievements has been made.However,for the development of visual navigation in practical applications,the existing systems still show obvious deficiencies in stability and robustness,which limit the promotion of visual navigation in engineering applications.Aiming at the robustness of visual navigation,this thesis deeply studies the reliability improvement measures in many aspects of visual navigation,and puts forward a series of novel methods of VSLAM and visual navigation.Experiments show that the proposed methods have higher robustness and accuracy than the existing related methods,and can better meet the needs of practical applications.Specifically,this thesis makes relevant contributions from the following perspectives of visual navigation1.For the back-end representation of environment modeling,this thesis proposes a multi-subgraph VSLAM back-end method.By using the pose-graph representation with multi-subgraphs,the environment model is divided into several independently traceable and aligned subgraphs.By using the independence of subgraph,the system can run continuously,and the fault tolerance rate against visual tracking failure is improved.At the same time,robust closed-loop constraints are introduced to eliminate redundant false topological associations in the case of false topological relationships between subgraphs This research focuses on the back-end of VSLAM system in visual navigation to improve the robustness of the system.2.In the front-end of VSLAM,where the data association is achieved,a dynamic keyframe selection strategy with adaptive adjustment is proposed.By introducing the mature cybernetics,the fixed threshold in the traditional keyframe selection method is replaced by the dynamic threshold.By using the PD controller as the adaptive implementation unit and the correlation strength of visual tracking status as the feedback of the controller,the adaptive adjustment of keyframe selection conditions is realized.By introducing dynamic adaptive term,the threshold parameter-adjustment workload of key frame selection is reduced,and the collected set of keyframe is optimized to improve the system accuracy and robustness.From the aspect of front-end in VSLAM system,the stability of visual tracking is improved.3.Based on the studies of the front-end and back-end in VSLAM mentioned above,and regarding the VSLAM system as the core,this thesis designs and constructs a visual environment exploration system with fault tolerance design.Considering the fragility of visual tracking,a ground segmentation algorithm based on deep neural network is used to realize the separation of mobile robot control from visual global localization,which enables the system to update the image input at any time and maintains reliable operation of each vision-based module.Secondly,an evaluation mechanism of the established environment model is designed,which is used to evaluate the completeness of the VSLAM model.This mechanism can automatically stop exploring at the right time,save the time of exploration,and ensure the feasibility of visual re-localization in the established model.The robust performance of visual navigation is improved by systematical architecture.4.Finally,in order to realize safe visual navigation,based on the system of visual exploration and visual modeling,this thesis proposes a belief evaluation method for safety path planning with visual localization in the established model.The potential confidence of visual localization in a region is evaluated by using the statistics of local keyframe density in the gridded environment and the global statistical analysis of keyframe distribution with a parametric distribution expression.Through this method,the robot can evaluate the success rate of visual localization of a planned arrival position before online visual navigation,so as to avoid the high-risk areas of visual tracking failure and achieve reliable visual navigation.