Abstract:Fast and efficient semantic segmentation of large-scale LiDAR point clouds is a fundamental problem in autonomous driving. To achieve this goal, the existing point-based methods mainly choose to adopt Random Sampling strategy to process large-scale point clouds. However, our quantative and qualitative studies have found that Random Sampling may be less suitable for the autonomous driving scenario, since the LiDAR points follow an uneven or even long-tailed distribution across the space, which prevents the model from capturing sufficient information from points in different distance ranges and reduces the model's learning capability. To alleviate this problem, we propose a new Polar Cylinder Balanced Random Sampling method that enables the downsampled point clouds to maintain a more balanced distribution and improve the segmentation performance under different spatial distributions. In addition, a sampling consistency loss is introduced to further improve the segmentation performance and reduce the model's variance under different sampling methods. Extensive experiments confirm that our approach produces excellent performance on both SemanticKITTI and SemanticPOSS benchmarks, achieving a 2.8% and 4.0% improvement, respectively. The source code is available at <a class="link-external link-https" href="https://github.com/huixiancheng/PCB-RandNet" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: in the autonomous driving scenario, the existing methods for processing large - scale LiDAR point clouds based on random sampling strategies cannot fully capture information at different distance ranges, resulting in a decline in the learning ability of the model, especially poor segmentation performance at medium and long - distance ranges. Specifically: 1. **Limitations of existing methods**: - Existing point cloud processing methods mainly adopt the Random Sampling (RS) strategy to process large - scale point clouds. However, the spatial distribution of LiDAR point clouds is usually uneven and even presents a long - tailed distribution, that is, the point cloud density in the close - range area is high, while the point cloud density in the long - distance area is low. - This uneven distribution makes the random sampling method sample fewer points in the medium and long - distance ranges, causing the model to be unable to fully learn the features of these areas, thus affecting the segmentation performance. 2. **Specific manifestations of the problem**: - The random sampling method retains more points in the close - range area, but discards many points in the medium and long - distance areas, resulting in a significant decline in the model's segmentation performance in these areas. - For example, Figure 1(d) shows that as the distance from the sensor increases, the model's segmentation performance drops sharply. To solve these problems, the author proposes a new sampling method - Polar Cylinder Balanced Random Sampling (PCB - RS), and further improves the stability and segmentation performance of the model by introducing the Sampling Consistency Loss (SCL). ### Main contributions 1. **Proposing a new sampling method PCB - RS**: By dividing the point cloud into different cylindrical blocks and performing balanced random sampling within each block, the down - sampled point cloud maintains a more balanced distribution at different distance ranges, thereby guiding the segmentation model to better learn the point distribution characteristics at different distance ranges. 2. **Introducing the sampling consistency loss SCL**: Reducing the model differences under different sampling methods and enhancing the model's robustness to different sampling methods. 3. **Experimental verification**: Experiments were carried out on two large - scale outdoor autonomous driving scene datasets, SemanticKITTI and SemanticPOSS, and the results show that this method significantly improves the performance of the baseline model. Through these improvements, the paper aims to improve the efficiency and accuracy of large - scale LiDAR point cloud semantic segmentation tasks in autonomous driving scenarios.

PCB-RandNet: Rethinking Random Sampling for LIDAR Semantic Segmentation in Autonomous Driving Scene

LiDAR-based Panoptic Segmentation via Dynamic Shifting Network

LiDAR Panoptic Segmentation for Autonomous Driving

LiDAR-based 4D Panoptic Segmentation via Dynamic Shifting Network

Efficient LiDAR Odometry for Autonomous Driving.

LiDAR Point Clouds Semantic Segmentation in Autonomous Driving Based on Asymmetrical Convolution

RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds

SegNet4D: Efficient Instance-Aware 4D Semantic Segmentation for LiDAR Point Cloud

Location-Guided LiDAR-Based Panoptic Segmentation for Autonomous Driving.

Panoptic-PolarNet: Proposal-free LiDAR Point Cloud Panoptic Segmentation

SyS3DS: Systematic Sampling of Large-Scale LiDAR Point Clouds for Semantic Segmentation in Forestry Robotics

DGPolarNet: Dynamic Graph Convolution Network for LiDAR Point Cloud Semantic Segmentation on Polar BEV

Hybrid CNN-LSTM Architecture for LiDAR Point Clouds Semantic Segmentation

PANet: LiDAR Panoptic Segmentation with Sparse Instance Proposal and Aggregation

SFPNet: Sparse Focal Point Network for Semantic Segmentation on General LiDAR Point Clouds

CPSeg: Cluster-free Panoptic Segmentation of 3D LiDAR Point Clouds

Epointda: an End-to-End Simulation-to-Real Domain Adaptation Framework for LiDAR Point Cloud Segmentation

LiDAR-Based Real-Time Panoptic Segmentation via Spatiotemporal Sequential Data Fusion

An Improved RandLa-Net Algorithm Incorporated with NDT for Automatic Classification and Extraction of Raw Point Cloud Data

Scan-based Semantic Segmentation of LiDAR Point Clouds: An Experimental Study