Study On The Deep-learning-based Real-time Registration And Place Recognition Algorithm For LiDAR Point Cloud

Real-time acquisition of the precise position and posture is the basis and prerequisite for a mobile robot to realize high-level applications,such as autonomous navigation or simultaneous localization and mapping.To solve the problems that existed when using the global positioning system(GPS)or internal measurement unit(IMU)in real-world scenes(such as signal loss or error accumulation),researchers started to focus on using the data acquired by environmental sensors equipped on the robot to calculate the position and posture.Among these sensors,Li DAR has the advantages of long-term sensing,high accuracy,and robustness to light conditions,making it the primary sensor for automatic vehicles and robots to perceive the threedimensional(3D)world.As the key algorithms for Li DAR to acquire position and posture,study on robust point cloud registration and place recognition algorithms in real-world,large-scale scenes has great significance for robots to realize autonomous navigation and expand application scenarios.This thesis focuses on designing real-time and robust point cloud registration and place recognition algorithms under the influence of large-scale scene point cloud,using 3D deep learning as the primary technical method.The main contents and achievements are as follows:(1)Aiming at the influences of the large-scale low-overlapping point cloud on the registration process,this thesis proposed a deep neural network named DDRNet,which can provide real-time and robust registration results for large-scale scenes.Unlike most registration methods that need to detect and match key points,DDRNet can directly sense the posture and predict registration results.This thesis designed multiple modules to realize the efficiency and robustness of the registration process,including the localspatial-aware feature encoder,the attentive-weighting module,and the pyramid feature decoder.We comprehensively validated DDRNet’s efficiency and robustness and compared it with the state-of-the-art registration methods using data acquired in various scenarios(including outdoor scenes,indoor scenes,and object point clouds).The results demonstrate that DDRNet can provide real-time and accurate registration results;it is also more robust than state-of-the-art methods to the variations in point density,noise,and overlapping ratios.The experiments on low-overlapping and no-overlapping scenarios show that DDRNet can deduce the registration results even though there is no overlap between the two scenes.(2)To solve the influences caused by complicated indoor scenarios,this thesis proposed a deep neural network,named Fore-net,to efficiently filter out outlier points(non-overlapping points and random outlier noises)and retain only inlier points,which can promote the performance of subsequent registration algorithms by optimizing the quality of source data.Fore-Net contains three modules to realize the efficiency and robustness of the filtering process,including the SPV-Conv-based feature descriptor,the dual-space-attention feature enhancement module,and the feature decoder.We comprehensively evaluate Fore-Net using various indoor scenarios;the results show that Fore-Net is more accurate than the state-of-the-art approaches and can stay robust under noise and point sparsity variance.The results further validated that by using ForeNet as a pre-processing step,the performance of registration algorithms,whether classical methods or learning-based methods,can be promoted in the low-overlapping scenarios.(3)Aiming at solving the place recognition problem in large-scale scenes,we proposed a place recognition network,EPR-Net,to efficiently encode global features for large-scale scenarios and realize real-time localization.Unlike existing methods that learn dense per-point features in the network,EPR-Net proved that using sparse features learned from local regions is also effective for constructing the place recognition network.We designed the feature encoder based on the random sampling method to learn local features efficiently and combine the graph convolutional network with the dual-space feature enhancement module to reinforce the network’s understanding of the whole scenario.We conducted experiments in various scenarios and compared EPRNet with state-of-the-art methods to validate the performance.The results show that EPR-Net can achieve accurate place recognition and localization,and our network is far more robust than the existing methods under variances of noise and point sparsity,making EPR-Net more suitable for use in the real-world environment.(4)To solve the problem that the supervised learning network relies on welllabeled training data,which reduces the practicability of our proposed network,we designed various automatic data labeling strategies.The proposed strategies enable the proposed networks(DDRNet and Fore-Net)can be trained and tested in a selfsupervised manner.They can further use hyperparameters to precisely control the scene properties such as overlapping ratio,point density,or noise intensity,enabling the network to be trained using scenes with more diversity.