基于知识蒸馏和定位引导的Pointpillars点云检测网络  

Pointpillars point cloud detection network based on knowledge distillation and location guidance

作　　者：赵晶 李少博 郭杰龙 俞辉 张剑锋 李杰 ZHAO Jing;LI Shaobo;GUO Jielong;YU Hui;ZHANG Jianfeng;LI Jie(School of Electrical Engineering and Automation,Xiamen University of Technology,Xiamen 361024,China;Fujian Institute of Research on the Structure of Matter,Chinese Academy of Sciences,Fuzhou 350108,China;Quanzhou Institute of Equipment Manufacturing,Haixi Institutes,Chinese Academy of Sciences,Quanzhou 362000,China;Xiamen Key Laboratory of Frontier Electric Power Equipment and Intelligent Controly,Xiamen 361024,China)

机构地区：[1]厦门理工学院电气工程与自动化学院,福建厦门361024 [2]中国科学院福建物质结构研究所,福建福州350108 [3]中国科学院海西研究院泉州装备制造研究中心,福建泉州362000 [4]厦门市高端电力装备及智能控制重点实验室,福建厦门361024

出　　处：《液晶与显示》2024年第1期79-88,共10页Chinese Journal of Liquid Crystals and Displays

基　　金：福建省科技计划(No.2021T3003);泉州市科技计划(No.2021C065L);福建省科技厅自然科学基金(No.2020J01285,No.2022J05285)。

摘　　要：激光雷达数据由于其几何特性,被广泛应用于三维目标检测任务中。由于点云数据的稀疏性和不规则性,难以实现特征提取的质量和推理速度间的平衡。本文提出一种基于体柱特征编码的三维目标检测算法,以Pointpillars网络为基础,设计Teacher-Student模型框架对回归框尺度进行蒸馏,增加蒸馏损失,优化训练网络模型,提升特征提取的质量。为进一步提高模型检测效果,设计定位引导分类项,增加分类预测和回归预测之间的相关性,提高物体识别准确率。本网络所做改进没有引入额外的网络嵌入。算法在KITTI数据集上的实验结果表明,相比于基准网络,在三维模式下的平均精度值从60.65%提升到了64.69%,鸟瞰图模式下的平均精度值从67.74%提升到70.24%。模型推理速度为45 FPS,在提升检测精度的同时满足了实时性要求。Lidar data is widely used in 3D target detection tasks due to its geometric characteristics.Due to the sparsity and irregularity of point cloud data,it is difficult to achieve the balance between the quality of feature extraction and the speed of reasoning.In this paper,a three-dimensional target detection algorithm based on body-column feature coding is proposed.Based on Pointpillars network,the Teacher-Student model framework is designed to distill the regression frame scale,increase distillation loss,optimize the training network model,and improve the quality of feature extraction.In order to further improve the model detection effect,the positioning guidance classification item is designed to increase the correlation between classification prediction and regression prediction,and improve the object recognition accuracy.The improvement of this network does not introduce additional network embedding.The experimental results of the algorithm on the KITTI dataset show that the average accuracy of the reference network in 3D mode is improved from 60.65%to 64.69%,and the average accuracy of the aerial view mode is improved from 67.74%to 70.24%.The model reasoning speed is 45 FPS,which meets the real-time requirements while improving the detection accuracy.

关 键 词：激光点云 三维目标检测 知识蒸馏 分类置信度 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]