Research On Semantic Map And Its Key Technology

In the context of the rapid development of deep learning,semantic maps have become a hotspot in the field of SLAM(Simultaneous Localization and Mapping),and have attracted the attention of a large number of researchers.To realize the understanding of the surrounding environment and objects,the semantic map is applied to the SLAM mapping method through semantic segmentation based on neural network,object detection,instance segmentation and other technologies.The difference between this method and the mainstream visual SLAM method is that it does not estimate the motion posture and environment of the camera based on the feature points and the underlying pixel level,but assists the mapping by using the semantic information in the environmental objects.Compared with the traditional SLAM mapping method,this method is more in line with the principle of human visual system.In addition,with the popularization of product-level deep acquisition equipment,it provides physical technical support for object detection and object semantic segmentation algorithms under visible light environment conditions,and provides performance advantages for building environmental semantic maps and object recognition algorithms.This paper analyzes the problem of how to construct a stable and effective environmental semantic map from three levels: understanding environmental semantic information,identifying environmental object information,and constructing a robust dynamic semantic map.In this paper,we research on multiple issues such as environmental semantic information recognition,topology node recognition,small sample object recognition,semantic map construction in a dynamic environment,and environmental object database construction.The main contributions of this article include:1.We propose a method for constructing a semantic map of the topological environment.This method recognizes the semantic information in the scene through the convolutional neural network,using its projection information to construct a 2D environment semantic map,and the RGB-D SLAM algorithm to construct a threedimensional map of the environment,and identifying topological node information by identifying fork nodes in the map.Finally,we construct a semantic map containing 2D environment semantic information,3D point cloud information and topology node information.2.Based on the Few-shot learning framework,we propose a network structure called EACM(embedded adaptive cross-modulation)mechanism and use it to solve the Few-shot learning.On the one hand,this method adjusts the cosine metric distance between different categories adaptively in the metric space,thereby improving the distinguishability and improving the classification effect.On the other hand,through a cross-modulation method between different layers of the neural network,the feature interaction between the support set and the verification set is enhanced,and the feature expression of the training samples is enriched,thereby improving the classification effect.In addition,we also use a neural network based on an adaptive attention mechanism to enhance the comparison of distance measures of semantically related categories in the embedded feature space,and suppress the comparison of distance measures of semantically unrelated categories.We multiply the measurement result of the cosine distance between all feature vectors by the relationship coefficient obtained through attention network training,so that the classification effect between different categories is more obvious.3.At present,most researches on semantic maps are based on the static environment as the premise of map construction,and often focus on the construction of 3D point cloud maps or the construction of camera trajectories,or on the recognition of environmental objects.In this paper,we use the dynamic point detection algorithm to filter the dynamic key points in the image,and combine the existing visual SLAM system with the semantic segmentation neural network and other algorithms to achieve the effect of building a dynamic semantic map.While constructing the dynamic semantic map,the mapping effect of the local map is optimized through the semantic information,and the environmental object information database is constructed.