【摘要】 自动驾驶与道路智能感知的发展对点云检测提出了新要求。路侧激光雷达点云不规则点构成的多变形态使其处理极具挑战性。传统点云分析方法依赖复杂度递增的局部几何提取器捕获三维信息，但此类方法加剧了算法复杂性与计算延迟，且近年性能趋于瓶颈。本研究认为在简单结构中增强深层网络对浅层特征的利用率是提升网络特征提取能力的有效途径，提出点密集连接网络，一种不依赖复杂提取器的通用点云处理架构。通过残差结构与密集连接模式，在深层聚合浅层多尺度语义特征，同时保持多数据集竞争力。在自制的激光雷达点云数据集上进行了对比实验，实验结果表明，本算法在保持轻量化的同时实现了优异的性能，能对常见道路目标实现精准分类。更多还原

【Abstract】 The development of autonomous driving and intelligent road perception has placed new demands on point cloud detection. The irregular and highly variable patterns formed by roadside LiDAR point clouds make them particularly challenging to process. Traditional point cloud analysis methods rely on increasingly complex local geometric extractors to capture 3D information, but these approaches exacerbate algorithmic complexity and computational latency, and in recent years their performance has plateaued. In this paper, we argue that enhancing a deep network’s utilization of shallow features within a simple architecture is an effective way to improve feature-extraction capabilities. We therefore propose the Point-Dense Connectivity Network, a general point cloud processing framework that does not depend on complex extractors. By combining residual structures with dense connections, it aggregates multi-scale shallow semantic features in its deeper layers while remaining competitive across multiple datasets. Comparative experiments on a custom LiDAR point cloud dataset demonstrate that our method achieves outstanding performance in precise classification of common road objects, all while maintaining a lightweight design.更多还原