Research Of Semantic Segmentation For LiDAR Point Clouds In Autonomous Driving

With the rapid development of artificial intelligence,autonomous driving is gradually becoming possible.To ensure adequate safety,autonomous driving systems need to achieve precise and robust scene perception.Therefore,Li DAR,which can accurately capture the 3D information of the surrounding environment,has become an indispensable sensor device for autonomous vehicles.The semantic segmentation task of the Li DAR point cloud is an essential basis for scene perception,which predicts the point-by-point semantic labeling of the 3D point cloud data collected by Li DAR,thus providing finegrained semantic perception information of the scene.This thesis researches and explores some problems of semantic segmentation of Li DAR point clouds in autonomous driving scenarios,and the main research results are as follows:1)To address the problem of the limited perceptual field of segmentation networks with 2D CNN as the backbone in projection-based semantic segmentation methods for point clouds,this thesis proposes a Transformer-based Li DAR semantic segmentation algorithm.By introducing the Vision Transformer(Vi T)structure to the semantic segmentation of range images,better local and global contextual information is learned and stronger long-range dependencies are obtained.Constructs a Balanced Non-Square Transformer module that uses a carefully designed Transformer based on a non-square shift window to capture feature dependencies at long local ranges.Experiments have shown that the proposed method achieves highly competitive performance on several datasets.2)Aiming at the problem that Li DAR point cloud semantic segmentation methods with large parameters or complex model structures cannot achieve speed and accuracy trade-off well,this thesis proposes a concise and efficient segmentation method.Through a rational convolutional kernel design,the introduction of a stronger non-linear activation function,and an auxiliary segmentation head that enhances the decoder features,the feature learning capability of the network is improved without introducing parameters and inference costs.Experiments show that the proposed method achieves state-of-theart performance on multiple datasets while being structurally simple and maintaining high inference speed.3)For the problem that the distribution density of Li DAR point clouds varies significantly at different distances,this thesis argues that random sampling does not apply to this typical long-tailed distribution scenario.This thesis proposes a new polar cylinder balanced random sampling method to maintain a more balanced distribution of the downsampled point cloud.Further,sampling consistency loss is also proposed to guide the model to learn rich feature representations under the two different sampling methods and to reduce the differences in model performance under the different sampling methods.Experiments show that both the proposed polar cylinder balanced random sampling method and the sampling consistency loss can improve the segmentation performance of the model at different distance ranges.In summary,this thesis presents innovative and effective solutions in three aspects for the core task of semantic segmentation of Li DAR point clouds for practical application scenarios in autonomous driving scenarios.Whether it is a new Transformer-based network structure,a new model that can balance speed and accuracy,or a new sampling method that can improve point cloud segmentation at medium and long distances,all shed light on research related to semantic segmentation of Li DAR point clouds,which is conducive to the continued progress and widespread use of autonomous driving technology.