Research And Application Of 3D Semantic Map Construction Based On Deep Learning

Intelligent robot is an important tool for humans to realize industrialization and information society.Understanding the surrounding environment and completing specific tasks autonomously according to the surrounding environment is an important symbol of robot intelligence.Synchronous localization and mapping(SLAM)technology is It is of great significance to achieve this goal.The existing SLAM algorithm realizes robot positioning by establishing a geometric map,and the geometric map can only ensure that the robot can avoid obstacles to reach a specified location,and cannot complete more advanced tasks.Therefore,this thesis studies a three-dimensional semantic map construction algorithm based on deep learning to improve the information level of the three-dimensional map and help the robot to better understand the surrounding environment.The main research contents of this article are as follows:1.An improved algorithm for deep matrix estimation based on unsupervised learning is proposed,which makes monocular vision integrate the advantages of binocular vision in the field of distance measurement,and at the same time enables the algorithm to flexibly adapt to new scenes.In this thesis,the source and target image pairs with known geometric relationships are input to the encoder and decoder respectively,and the required model is finally obtained.The process is mainly divided into three steps:first,the source image is input to the encoder,and the depth matrix prediction model is obtained through training;the second step is to fuse the target image and the depth matrix at the decoder to obtain the reconstructed image;the third step is to source The photometric error of the image and the reconstructed image is used as the reconstruction error of the automatic encoder to train the network.2.Integrate the object semantic information classifier based on deep learning into the ORB-SLAM system,merge the semantic information with a single geometric map,improve the information level of the traditional three-dimensional map,and in order to ensure the real-time operation capability of the system,introduce A strategy for extracting important information from key frames,and using conditional random fields to optimize and update the extracted semantic information of objects,improve the accuracy of detection.In the tracking and positioning process,the extracted sequence semantic information is fused and an object in a three-dimensional environment is generated.The object contains information about the category,the three-dimensional model in space,and the world coordinate position.At the same time,the relationship between the object and the key frame is established Object semantic information is fused into SLAM system,and finally based on fast rasterized multi-thread octree map representation algorithm,which improves the access efficiency of 3D semantic map.3.Program and implement the improved algorithm proposed in this thesis,optimize the algorithm as a whole,reduce the calculation amount of the front and back ends,and choose to test on the standard data set KITTI,TUM,through the establishment of experimental equipment and implement the entire algorithm Process.Experimental results show that,in some application scenarios,the proposed three-dimensional semantic map construction algorithm improves certain accuracy compared with traditional algorithms,and meets real-time requirements to a certain extent.