Research On Construction Technology Of Topology-Metric Hybrid Map Based On Graph Neural Network

Environment modeling is a key capability for unmanned ground vehicles.The current mainstream method of constructing high-precision maps based on Li DAR SLAM technology will face high maintenance costs in the dynamic combat environment,and lack sufficient robustness and ability to understand and recognize scene content.Aiming at the road scene modeling problem of unmanned ground vehicles in the urban combat environment,this paper studies the real-time construction method of topology-metric hybrid map based on graph convolutional neural network.The main research results are as follows:1.An intersection classification algorithm based on graph neural network(GC-ICNet)is proposed.The algorithm introduces the idea of graph classification.First,the KNN and Query Ball methods are used to realize the graph structure representation of point cloud data,and then the graph readout mechanism and feature skip connection structure are designed,and finally an end-to-end intersection classification model is constructed.The real vehicle test confirms the effectiveness of the algorithm on the intersection classification task in the urban road scene.2.A road segmentation model based on graph neural network(FK3D-RSNet)is designed.The model integrates knowledge-driven manual features and data-driven graph attention convolution features.First,a graph attention convolution mechanism that fuses point cloud local neighborhood spatial distance and feature distance is introduced.According to the characteristics of the road region in urban road scenes,a manual feature extraction module is designed,and finally an end-to-end road segmentation framework is realized by fusing manual features and graph attention convolution features.On the basis of solving the road segmentation problem,the model further considers the possible smoke interference that unmanned ground vehicles may face in the urban combat environment,and is successfully extended to the task of smoke segmentation.The real vehicle test results show that the model can effectively solve the road segmentation problem under the interference of smoke in urban road scenes.3.Based on the above research results,a set of real-time construction and presentation system(T-MHMapping)of topology-metric hybrid map for road network environment is developed.While constructing the topology map,the system integrates a variety of metric information in the road network environment,and realizes the ability to understand and identify the content of the road scene.It can complete the real-time construction of the topology-metric hybrid map in the dynamic and harsh environment with the lack of GNSS signal and the presence of smoke interference.It helps to improve the ability of unmanned ground vehicles to model the road scene in the urban combat environment,and then complete the tasks such as planning and navigation.