面向道路交通场景的高效3D目标检测  

High-efficiency 3D object detection for road traffic scenes

作　　者：陆军[1] 鲁林超 翟晓阳 刘霜 LU Jun;LU Linchao;ZHAI Xiaoyang;LIU Shuang(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学智能科学与工程学院,黑龙江哈尔滨150001

出　　处：《智能系统学报》2025年第1期91-100,共10页CAAI Transactions on Intelligent Systems

基　　金：黑龙江省自然科学基金项目(F201123)。

摘　　要：针对当前两阶段的点云目标检测算法PointRCNN:3D object proposal generation and detection from point cloud在点云降采样阶段时间开销大以及低效性的问题,本研究基于PointRCNN网络提出RandLA-RCNN(random sampling and an effectivelocal feature aggregator with region-based convolu-tional neural networks)架构。首先,利用随机采样方法在处理庞大点云数据时的高效性,对大场景点云数据进行下采样;然后,通过对输入点云的每个近邻点的空间位置编码,有效提高从每个点的邻域提取局部特征的能力,并利用基于注意力机制的池化规则聚合局部特征向量,获取全局特征;最后使用由多个局部空间编码单元和注意力池化单元叠加形成的扩展残差模块,来进一步增强每个点的全局特征,避免关键点信息丢失。实验结果表明,该检测算法在保留PointRCNN网络对3D目标的检测优势的同时,相比PointRCNN检测速度提升近两倍,达到16 f/s的推理速度。Based on the 3D object proposal generation and detection from pointcloud,namely PointRCNN network,this study proposes an RandLA-RCNN architecture to address the issues of high time cost and inefficiency in the point cloud downsampling stage of the current two-stage point cloud object detection algorithm.Firstly,by taking advantage of the efficiency of random sampling method,the large-scale point cloud data are downsampled to handle massive point cloud data.Then,the spatial positions of each neighboring point of the input point cloud are encoded to effectively enhance the ability of each point to extract local features from the neighborhood.Attention-based pooling rules are used to aggregate local feature vectors and obtain global features.Finally,an extended residual module formed by stacking multiple local spatial encoding units and attention pooling units is used to further enhance the global features of each point and avoid the loss of key point information.Experimental results show that this detection algorithm retains the advantages of PointRCNN network in detecting 3D objects,while achieves nearly twice the detection speed compared with PointRCNN,reaching an inference speed of 16 frames per second.

关 键 词：深度学习 3D目标检测 点云 随机采样 局部特征聚合 注意力机制 自动驾驶 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]