Research On Semantic Map Construction Algorithm Of Mobile Robot

With the development of science,technology and economy,intelligent robot has gradually become the focus and trend of robot research.In the process of working,mobile robot not only needs to realize its own positioning,but also needs to build the current map.On this basis,if we can build a map with practical meaning that reflects the key objectives and regions in the actual scene,it can help robots better understand the actual situation in the scene.Based on visual slam and deep neural network,this paper constructs a semantic map that reflects multiple targets and regions in the actual scene.The main work of this paper is as follows:1.Research on target construction algorithm of semantic map.Based on the target detection network,the algorithm is improved.IoU and appearance matching are comprehensively used to calculate the similarity score,so as to realize the tracking of the same target in continuous frames and avoid repeated detection and error detection.Because the proportion of pedestrians and vehicles in the detection boundary box is different,the pedestrian foothold method is used to calculate the pedestrian position information.2.Research on region construction algorithm of semantic map.On the basis of semantic segmentation network,this paper focuses on the respective characteristics of trees and building areas.C omprehensively compare different clustering algorithms to find an algorithm suitable for tree region extraction.The analysis shows that the building area is different from the characteristics of trees.It is concluded that the plane fitting algorithm is more suitable for extracting the building area,and the algorithm is designed.3.Research on slam pose optimization algorithm based on semantic plane constraints.With the help of the above semantic segmentation Neural Network,the ground is detected first.The system with plane information is used for initialization and PNP tracking,and the re projection consistency is used to expand the plane area,which is optimized in the sliding window with plane information.It helps the slam system to achieve better pose estimation and makes the final semantic map more accurate.4.System experiment and analysis.The feasibility and positioning accuracy of the proposed algorithm are verified by using the public data set and the actual scene data set taken by binocular camera.