Research On Large-scale Mapping Of Stereo Vision Mobile Robot

With the rapid development of science and technology,mobile robots with automatic navigation functions have entered various fields such as inspection,rescue and security.In order to realize the automatic navigation of the mobile robot,it is necessary to establish a navigation map for the mobile robot in advance.At present,lidar sensors are mainly used for building navigation maps for mobile robots.At present,lidar sensors are mainly used for building navigation maps for mobile robots.However,lidar sensors are expensive,it is easy to produce wrong measurements when they encounter non-reflective or black objects,which affects the accuracy of positioning and mapping.Therefore,the use of visual sensors to complete positioning and mapping has gradually become an important research direction.At present,localization and mapping algorithms based on visual sensors can achieve good results indoors,but there are problems such as low efficiency,poor accuracy,and weak robustness in outdoor large-scale environments.Generally,mobile robots need to satisfy autonomous movement in outdoor large-scale environments,this paper will build a large-scale mapping algorithm for mobile robots In this paper,large-scale is relative to the small-scale indoor closed environment.Due to the complex environment(Weak texture and under-illuminated areas)and large amount of information in large-scale environments,the depth estimation error increase in square terms with depth distance,long-distance reconstruction error is large,especially modeling near the edge of the object is easy to diverge,generating more noise,which ultimately affects the accuracy and robustness of large-scale mapping algorithms.In order to reduce the depth estimation error in the traditional large-scale mapping algorithm,this paper proposes an improved inter-group relationship to construct a cost-volume deep learning architecture.The network consists of a feature extraction module,a cross-form spatial pyramid module,and three-dimensional feature matching Composed of fusion modules.Experiments on the KITTI dataset based on the stereo matching algorithm after the network model training show that the stereo matching algorithm has a certain improvement in matching accuracy and efficiency compared with the previous stereo matching algorithm.Although the stereo matching algorithm based on deep learning can improve the accuracy and efficiency of stereo matching,when the actual large-scale scene and the training image scene have a low similarity,the error probability of depth estimation is large.In order to make up for this defect,this paper designs a depth estimation fusion module,which fuses the traditional depth estimation module with the deep learning-based depth estimation module.The depth image selection module is designed based on the principle of optical consistency to improve the accuracy of large-scale mapping algorithms.Large-scale mapping algorithms are tested on KITTI data sets and real scenes.In the KITTI 03,05,07,08 sequence data set,it is verified that the large-scale mapping algorithm in this paper meets the accuracy requirements.Carry out large-scale mapping experiments in the closed-loop,weak texture and weak light areas of real scenes.The experimental results show that the large-scale mapping algorithm in this paper can also ensure the accuracy of the trajec tory of the mapping in the real scene area,and the algorithm has strong robustness.