Application Of Deep Learning Technology In LiDAR Place Recognition Task

Real-time and accurate acquisition of the position of intelligent mobile robots in the map is the basis for realizing tasks such as autonomous robot navigation and path planning.To solve the problems of GPS in scenarios such as signal interference and sheltered space where signals are easily lost and inertial navigation systems with initialization requirements and error accumulation,Researchers have begun to focus extensively on how robots can use their own mounted environment-aware sensors for accurate positioning.Li DAR,as an environment-aware sensor,can show greater adaptability to changes in ambient light and is suitable for working in real-world environments with long time spans.The point cloud place recognition algorithm is the key for intelligent robots to obtain their own position in the global map.The place recognition algorithm needs to describe the features of point clouds,and the traditional artificially designed features are not robust under the influence of point cloud density changes and point cloud noise.With the development of deep learning technology,point cloud deep learning technique shows far more effect than traditional point cloud processing and shows its powerful ability in several tasks.In this thesis,we will focus on the accuracy and robustness of place recognition algorithms in point cloud noise and low-density point cloud environments,etc.The main research contents and results of the deep learning techniques are as follows:(1)A summary of point cloud deep learning technology is summarized,including the process from perceptron to neural network,the role of each layer of convolutional neural network and common methods,as well as the types of neural network optimization algorithms and their advantages and disadvantages,etc.The summary of the theoretical part provides the basis for the subsequent research.(2)An efficient and robust place recognition network under point cloud noise and low density environment is proposed.The network is a learning-based network consisting of three main modules: a feature encoding module that combines point-based and voxel-based structures to learn local features efficiently and effectively;an attention weighting module that aggregates spatial relationships and reinforces features from a global perspective;and a feature fusion module for generating global The feature fusion module is used to generate global descriptors for feature retrieval.Experimental results show that it can accurately retrieve and match scene with a running time of 9ms per point cloud frame,and is more robust to changes in point density and noise intensity.The network’s efficient,robust and generalization capability makes it more suitable for application in complex environments.(3)To address the problem that the datasets for training and testing of the proposed place recognition network are open source outdoor datasets,an indoor point cloud dataset is constructed to test the generalization of the network,and the network parameters obtained from training on the open source dataset are applied to the pre-processed point cloud data.The experimental results show that the network can still achieve an Top@5 average recall of more than 80% in the complex and highly repetitive indoor scenes.