Research On Deep Learning Based Place Recogntion Of Large-scale Lidar Point Cloud

Transportation has shifted from traditional fuel-fueled vehicles to new-energy vehicles and is developing towards sharing and intelligence.Especially in recent years,the development of artificial intelligence technology based on deep learning has driven the development and landing of autonomous vehicles.Deep learning algorithms have demonstrated powerful performance and generalization capabilities in areas such as image and natural language processing,which provides a powerful solution for the perception and positioning system of autonomous vehicles.At the same time,the development of lidar has given intelligent mobile robots,including self-driving car systems,the ability to directly sense the three-dimensional environment,with a farther sensing distance,more laser beams,and higher resolution.This improves the intelligence of the perception and positioning system but also brings more technical challenges.The perception positioning system of intelligent mobile robots such as self-driving cars not only needs to detect objects(e.g.,vehicles,pedestrians)in the surroundings but also requires accurate self-positioning.Simultaneous localization and mapping(SLAM)is one of the primary and critical technologies for autonomous navigation and positioning of mobile robots.Among them,loop closure detection and relocalization are the key technologies for correcting the mapping pose error and positioning application of the SLAM.The traditional methods are mainly relying on GPS for assistance but deeply affected by the obstruction of trees and buildings.The rise of deep learning in 3D point cloud research provides the possibility to solve this problem by relying on Li DAR point clouds only.And this paper studies the large-scale point cloud place recognition based on deep learning,which provides a powerful method to loop closure detection and relocalization in Li DAR slam.The main research contents of this paper are as follows:(1)Research on deep learning methods for extracting and encoding neighborhood structure features of sparse 3D point clouds.According to the characteristics and challenges of point clouds,study the pipeline for deep learning network for point clouds,and analyze the method and problem statement of point cloud-based place recognition in large-scale scenes.(2)Study the deep learning-based place recognition of large-scale Li DAR point clouds.This paper proposes a novel network LPD-Net(Large-scale Place Description Network),through adaptive neighborhood local geometric feature extraction and dynamic graph network aggregation,learning local structure features and context semantic features.Then generate the corresponding global descriptors through Net VLAD Layer.The training and evaluation on the Oxford Robot Car dataset prove that LPD-Net has achieved the SOTA(State-Of-The-Art)performance of point cloud-based place recognition Benchmark.(3)Research on point cloud compression reconstruction and feature latent space representation method of large-scale point clouds.This paper proposes a point cloud reconstruction network,LPD-AE(Large-scale Place Description-Auto Encoder),which uses a hierarchical generation strategy to reconstruct large-scale point clouds from latent space vectors,from coarse to fine.The bijection mapping between the point cloud and the global description vector establishes the latent space representation method of the point cloud,which can be used to replace the point cloud to achieve loop-closure detection,relocalization,and map compression storage and transmission task.Trained and evaluated on Oxford Robot Car,KITTI,and self-acquisition SEU-FX dataset,the quantitative and qualitative results show the adequate performance of point cloud reconstruction.(4)This paper designs a multi-sensor fusion perception platform for real autonomous driving applications,which integrates lidar,binocular vision,GPS / INS fusion positioning navigation system,and collects SEU-FX dataset,including the repeated data of the same path in urban and campus scene under different time,weather,and light conditions,which can be applied to long-term scene recognition,relocation and other related tasks.Real applications,including scene loop closure detection and relocalization,have confirmed the convincing performance of point cloud latent space representation in Li DAR SLAM,demonstrating its promise potency for application deployment in real autonomous driving.