结合通道剪枝和通道注意力的轻量型车辆点云补全网络  

Lightweight Vehicle Point Cloud Completion Network Combined with Channel Pruning and Channel Attention

作　　者：杨晓文[1,2,3] 冯泊栋 韩慧妍 况立群 韩燮[1,2,3] 何黎刚 YANG Xiaowen;FENG Bodong;HAN Huiyan;KUANG Liqun;HAN Xie;HE Ligang(School of Computer Science and Technology,North University of China,Taiyuan 030051,China;Shanxi Key Laboratory of Machine Vision and Virtual Reality,Taiyuan 030051,China;Shanxi Province’s Vision Information Processing and Intelligent Robot Engineering Research Center,Taiyuan 030051,China;Department of Computer Science,The University of Warwick,Coventry CV47AL,U.K.)

机构地区：[1]中北大学计算机科学与技术学院,太原030051 [2]机器视觉与虚拟现实山西省重点实验室,太原030051 [3]山西省视觉信息处理及智能机器人工程研究中心,太原030051 [4]华威大学计算机科学系,英国考文垂CV47AL

出　　处：《计算机工程与应用》2025年第1期232-242,共11页Computer Engineering and Applications

基　　金：国家自然科学基金(62272426);山西省科技重大专项计划“揭榜挂帅”项目(202201150401021);山西省自然科学基金(202203021212138,202303021211153);山西省科技成果转化引导专项(202104021301055)。

摘　　要：针对现有的点云补全网络多关注于补全的精度而忽视补全效率问题,提出了一种轻量型点云补全网络来准确、高效地修复自动驾驶中的不完整车辆点云。为了提高网络推理效率,采用一种高效的一次性通道剪枝技术提高网络的补全效率;在特征提取阶段,网络加入通道注意力模块,将加权特征与全局特征拼接,通过两层多维特征信息提取,得到最终的特征向量;将特征向量传入双解码器结构中,分别通过全连接层和多层感知机生成稠密的粗糙点云和输入点云偏差值;将粗糙点云与输入点云偏差值相加得到最终的精细化完整点云。在PCN数据集和KITTI数据集上进行实验,实验结果表明在补全缺失车辆信息的实时性上有着显著的提升,并且在补全精度上也有不错的表现。A lightweight point cloud completion network is proposed to accurately and efficiently repair incomplete vehicle point clouds in autonomous driving scenarios,addressing the efficiency issue often overlooked by existing point cloud completion networks that primarily focus on accuracy.Firstly,to enhance inference efficiency,an efficient one-shot channel pruning technique is employed to improve the completion speed of the network.Secondly,a channel attention module is integrated into the network during the feature extraction phase,which combines weighted features with global features.This dual-layered multidimensional feature extraction yields the final feature vector.Subsequently,the feature vector is fed into a dual decoder structure,generating dense coarse point clouds and input point cloud deviation values through fully connected layers and multi-layer perceptron,respectively.Finally,by adding the coarse point cloud and input point cloud deviation values,a refined and complete point cloud is obtained.Experimental evaluations conducted on the PCN dataset and KITTI dataset demonstrate significant improvements in real-time completion of missing vehicle information.Moreover,favorable results are achieved in terms of completion accuracy as well.

关 键 词：点云补全 通道剪枝 通道注意力 轻量型 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]