| In recent years,with the development of artificial intelligence,assisted driving technology has become a hot research field,and LiDAR simultaneous localization and mapping(LIDAR SLAM)is widely used as a solution to perceive the environment for assisted driving systems in unknown areas and special situations.However,there are two important problems in the application process of LiDAR SLAM in actual scenes: First,when LiDAR SLAM always calculates odometers in real time,it relies on the matching between point clouds,which is easily affected by moving objects in the scene.Second,the estimation of the current pose by the LiDAR SLAM system is based on the previous estimation results,so its error will continue to accumulate as the running time increases.Therefore,in response to these two problems,this thesis focuses on the research on the Transformer-based LiDAR point clouds moving object segmentation technology and the coded binary feature map-based loop closure detection technology,builds a data acquisition simulation platform and relevant visualization software to present algorithm demonstration results.The main work of this thesis is as follows.(1)Aiming at the problem that the conventional point clouds semantic segmentation framework is difficult to learn the dynamic features of objects in the scene,this thesis studied and implemented a Transformer-based LiDAR moving object segmentation method.A calculation method of residual feature map was introduced to improve the dynamic feature expression ability of the network.At the same time,the multi-head self-attention idea in Transformer was combined,and its network structure was improved,and the window selfattention calculation was used to replace the global self-attention to reduce the amount of calculation and improve learning efficiency.Then,the method in this thesis was compared and tested in a series of sets of scenarios.The experimental results showed that the method of LiDAR dynamic object segmentation based on Transformer in this thesis has higher accuracy than other methods.(2)In view of the problems of poor real-time performance,insufficient detection accuracy,and dependence on a large number of training samples in various existing loop closure detection methods,this thesis studied and implemented a loop closure detection method based on binary feature maps.The method extracted a binary feature map with rotation invariance as the descriptor of the scene by dividing and coding the voxels of the point cloud data in the scene and performing LoG-Gabor filtering under the Fourier transform.The scene similarity could be matched according to the Hamming distance easily.Then,the loop closure detection method in this thesis was compared and tested in multiple sets of scenarios.The experimental results shown that the loop closure detection method in this thesis could effectively improve the real-time and accuracy of loop closure detection without relying on sample data for training.(3)In order to further verify the performance of the lidar point cloud dynamic target segmentation and loop closure detection method proposed in this thesis,this thesis built an assisted driving simulation platform under laboratory conditions to collect lidar point cloud data,and designed and implemented the visualization software of the lidar point cloud dynamic target segmentation and loop closure detection method described in this thesis under the framework of PyQt5. |
