Research On Visual SLAM Based On Point-Line Features And Deep Learning

With the changes in industrial production and daily life needs,people put forward higher requirements for the intelligence of mobile robots.Simultaneous Localization and Mapping(SLAM),as a key technology for the intelligence of mobile robots,has developed rapidly in recent years.Due to the cheap price of visual sensors and the rich image information collected,Visual SLAM has become a hot field in computer vision.Most of the current state-of-the-art Visual SLAM is based on sparse point features.However,in low-texture environments,it is difficult for point featurebased SLAM to find enough point features for pose estimation.Therefore,this paper studies the application of line features in Visual SLAM.In addition,most of the existing Visual SLAM uses the closed-loop detection algorithm based on the Bag-ofWords model,which uses artificial features,resulting in low accuracy.To this end,this paper proposes a loop-closure detection algorithm based on EfficientNet-UMAP to improve the accuracy of loop-closure detection.The main research contents of the article are as follows:1.Aiming at the problem of information redundancy in the Visual SLAM based on point-line features in high-texture environments,a feature extraction strategy based on image entropy is proposed,using the image entropy of the input image to judge the texture richness of the input image,and according to the texture,richness selects appropriate parameters to extract point-line features.In addition,aiming at the problem of decreased positioning accuracy caused by the difficulty of line feature data association,an error model based on weighted ideas is proposed,and the previously acquired image entropy is used to divide the weights of points and lines,and a consistent error model is constructed.Finally,the algorithm in this paper is tested on the KITTI and Eu RoC datasets,and it is proved that the algorithm in this paper improves the accuracy and real-time performance of the system in a high-texture environment;in the low-texture environment,the real-time performance of the system is improved while ensuring the positioning accuracy and robustness of the system.2.Aiming at the poor accuracy of the traditional closed-loop detection algorithm based on the Bag-of-Words model,a closed-loop detection algorithm based on EfficientNet is proposed.First,the pre-trained model trained by EfficientNet on the ImageNet dataset is used to extract image features,and then the cosine similarity is used to perform feature matching to find closed loops.In addition,for the problem of redundant features in the feature vectors extracted by EfficientNet,a closed-loop detection algorithm based on EfficientNet-UMAP is proposed,which uses the UMAP dimensionality reduction algorithm to reduce the original features,and then uses cosine similarity to perform feature matching to find closed loops.Finally,the algorithm of this paper is tested on the datasets of New College and City Centre,which verifies the effectiveness of the method in this paper.