文献类型：期刊文献

英文题名：Pillar-Based Adaptive Sparse Transformer with Cost-Optimized Positive Sample Selection for 4D Radar Object Detection

作者：Chen, Tongzhou[1];Wu, Danfeng[2,3];Zhou, Fenfen[2,3];Jing, Hui[1];Kuang, Minchi[4];Zhang, Xueyan[2,3]

第一作者：Chen, Tongzhou

通讯作者：Wu, DF[1];Wu, DF[2]

机构：[1]Guilin Univ Elect Technol, Sch Mech & Elect Engn, Guilin, Peoples R China;[2]Beijing Union Univ, Beijing Key Lab Informat Serv Engn, Beijing, Peoples R China;[3]Beijing Union Univ, Coll Robot, Beijing, Peoples R China;[4]Tsinghua Univ, Dept Precis Instrument, Beijing, Peoples R China

第一机构：Guilin Univ Elect Technol, Sch Mech & Elect Engn, Guilin, Peoples R China

通讯机构：[1]corresponding author), Beijing Union Univ, Beijing Key Lab Informat Serv Engn, Beijing, Peoples R China;[2]corresponding author), Beijing Union Univ, Coll Robot, Beijing, Peoples R China.|[1141739]北京联合大学机器人学院;[11417]北京联合大学;[11417103]北京联合大学北京市信息服务工程重点实验室;

年份：2025

外文期刊名：INTERNATIONAL JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS RESEARCH

收录：EI(收录号：20254619499895);Scopus(收录号：2-s2.0-105021417514);WOS:【ESCI(收录号:WOS:001610528900001)】；

基金：This work was partly supported by School of Mechanical and Electrical Engineering at Guilin University of Electronic Technology, Beijing Key Laboratory of Information Service Engineering at Beijing Union University, College of Robotics at Beijing Union University, and Department of Precision Instrument at Tsinghua University.

语种：英文

外文关键词：4D imaging radar; Object detection; Transformer; Deep learning; Autonomous driving

摘要：The reliability and cost-effectiveness of 4D millimeter wave radar in adverse weather conditions make it an indispensable auxiliary sensor in autonomous driving. In response, we propose PSTOPS, a novel and effective 3D object detection framework based on 4D radar. Current 3D detectors based on 4D radar typically employ 3D convolutional backbones. However, because of their limited receptive fields, these backbones cannot efficiently capture large-scale contextual information, which is crucial for object detection. To address this issue, we adapt a single-step, window-based pillar transformer backbone. This backbone leverages the self-attention mechanism to achieve long-range relationships between pillars while effectively utilizing the sparsity of radar point clouds, naturally avoiding substantial computational overhead. Additionally, we found that during training, center-based label assignment often fails to generate sufficient positive samples, and anchor-based label allocation often suffers from imbalance when dealing with objects of different scales. To solve these problems, we designed a dynamic cross-label detection head. This head dynamically assigns positive samples for each object from a cross-shaped region, ensuring an adequate and balanced number of positive samples during training. Although only using 4D imaging radar, PSTOPS achieves performance comparable to the methods that fuse 4D radar and cameras.We evaluate the proposed method PSTOPS on the dataset View-of-Delft (VoD). Our method achieves 50.99% for the Entire Annotated Area and 71.55% for the Region of Interest. This demonstrates that 4D radar possesses significant potential for 3D object detection applications.