基于跨域时空特征匹配的路端3D目标检测无监督学习方法研究

doi:10.19562/j.chinasae.qcgc.2025.09.003

摘要/Abstract

摘要：

路端3D目标检测能提供广域交通参与者信息，可有效赋能单车自动驾驶。一般而言，在新的路口部署感知系统需要采集大量数据并进行人工标注以保证训练模型的检测精度，部署耗时长且成本高。针对上述问题，本文提出了一种面向路侧端的无监督域自适应算法，通过现有高质量标注路端数据与未知路端场景数据的跨域时空特征匹配，实现高效网络知识迁移与高准确率3D目标检测。首先，构建了考虑跨域训练数据分布均衡的源域训练模型（RoadPillars），能有效减少模型对原有数据分布的过拟合，提升模型泛化性。其次，设计了考虑连续序列信息空间一致性的跨域迁移框架，确保迁移过程稳定鲁棒。在3个不同路端场景和雷达类型的公开数据集展开的两组迁移测试结果显示，本文无监督域适应算法相较于目前最新方法的平均精度提升了24.2%和8.0%，有效增强了路端感知中目标检测模型的泛化性与可靠性。

关键词: 路端感知, 3D目标检测, 迁移学习, 无监督域适应, 特征匹配

Abstract:

Roadside 3D target detection provides wide-view traffic information， effectively enhancing single-vehicle autonomous driving. In general， deploying a perception system at a new intersection requires a large amount of data collection and manual labeling to ensure the detective accuracy of the training model， which is time-consuming and costly. For the above-mentioned problems， in this paper an unsupervised domain adaptation （UDA） algorithm is proposed for roadside application， which achieves efficient knowledge transfer and accurate 3D object detection by matching cross-domain spatiotemporal features between high-quality labeled roadside data and unknown roadside scenes. Firstly， a source-domain training model （RoadPillars） is constructed that balances cross-domain data distributions， effectively reducing overfitting of the model to the original data distribution and improving model generalization. Moreover， a cross-domain transfer scheme is designed that ensures spatial consistency of continuous sequences for stable and robust UDA. The experimental results on three distinct roadside scenes and LiDAR types from public datasets demonstrate that the unsupervised domain adaption algorithm proposed in this paper achieves an average accuracy improvement of 24.2% and 8.0% over state-of-the-art approach， significantly improving the generalization and reliability of roadside perception.

Key words: roadside perception, 3D object detection, transfer learning, unsupervised domain adaptation, feature matching

龚伟,王亚飞,汪博文,李泽星,孙家铭. 基于跨域时空特征匹配的路端3D目标检测无监督学习方法研究[J]. 汽车工程, 2025, 47(9): 1665-1673.

Wei Gong,Yafei Wang,Bowen Wang,Zexing Li,Jiaming Sun. Study on Unsupervised Learning Method for Roadside 3D Object Detection Based on Cross Domain Spatiotemporal Feature Matching[J]. Automotive Engineering, 2025, 47(9): 1665-1673.

图/表 14

图1

图2

图3

图4

图5

图6

图7

表1

表2

表3

图8

图9

图10

表4

参考文献 22

[1]	丁飞，张楠，李升波，等. 智能网联车路云协同系统架构与关键技术研究综述［J］. 自动化学报， 2022， 48（12）： 2863-2885.
	DING Fei， ZHANG Nan， LI Shengbo， et al. A survey of architecture and key technologies of intelligent connected vehicle-road-cloud cooperation system ［J］. Acta Automatica Sinica， 2022， 48（12）： 2863-2885.
[2]	张毅，裴华鑫，姚丹亚. 车路协同环境下车辆群体协同决策研究综述［J］. 交通运输工程学报， 2022， 22（3）： 1-18.
	ZHANG Yi， PEI Huaxin， YAO Danya. Research review on cooperative decision-making for vehicle swarms in vehicle-infrastructure cooperative environment［J］. Journal of Traffic and Transportation Engineering， 2022， 22（3）： 1-18.
[3]	YU Haibao， LUO Yizhen， SHU Mao， et al. Dair-v2x： a large-scale dataset for vehicle-infrastructure cooperative 3d object detection［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2022： 21361-21370.
[4]	YU Haibao， YANG Wenxian， RUAN Hongzhi， et al. V2x-seq： a large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2023： 5486-5495.
[5]	HAO Ruiyang， FAN Siqi， DAI Yingru， et al. Rcooper： a real-world large-scale dataset for roadside cooperative perception ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2024： 22347-22357.
[6]	QI C R， SU Hao， MO Kaichun， et al. PointNet： deep learning on point sets for 3D classification and segmentation ［C］. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition， 2017： 652-660.
[7]	QI C R， YI Li， SU Hao， et al. Pointnet++： deep hierarchical feature learning on point sets in a metric space ［J］. Advances in Neural Information Processing Systems， 2017， 30.
[8]	SHI Shaoshuai， WANG Xiaogang， LI Hongsheng. Pointrcnn： 3D object proposal generation and detection from point cloud［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2019： 770-779.
[9]	ZHOU Yin， TUZEL Oncel. Voxelnet： end-to-end learning for point cloud based 3D object detection ［C］. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition， 2018： 4490-4499.
[10]	LANG Alex H， VORA Sourabh， CAESAR Holger， et al. Pointpillars： fast encoders for object detection from point clouds ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2019： 12697-12705.
[11]	YAN Yan， MAO Yuxing， LI Bo. Second： sparsely embedded convolutional detection ［J］. Sensors， 2018， 18（10）： 3337.
[12]	WU Jianqing， XU Hao， SUN Yuan， et al. Automatic background filtering method for roadside LiDAR data ［J］. Transportation Research Record， 2018， 2672（45）： 106-114.
[13]	YANG Lei， YU Kaicheng， TANG Tao， et al. Bevheight： a robust framework for vision-based roadside 3D object detection ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2023： 21611-21620.
[14]	ZIMMER Walter， BIRKNER Joseph， BRUCKER Marcel， et al. Infradet3D： multi-modal 3D object detection based on roadside infrastructure camera and lidar sensors ［C］. 2023 IEEE Intelligent Vehicles Symposium （IV）， 2023： 1-8.
[15]	刘正发，吴亚，刘佩根，等. 基于特征和标签联合分布匹配的智能驾驶跨域自适应目标检测［J］. 汽车工程， 2023， 45（11）： 2082-2091.
	LIU Zhengda， WU Ya， LIU Peigen， et al. Cross-domain object detection for intelligent driving based on joint distribution matching of features and labels ［J］. Automotive Engineering， 2023， 45（11）： 2082-2091.
[16]	胡杰，徐博远，熊宗权，等. 基于多尺度掩码分类域自适应网络的跨域目标检测算法［J］. 汽车工程， 2022， 44（9）： 1327-1338.
	HU Jie， XU Boyuan， XIONG Zongquan， et al. Cross-domain object detection algorithm based on multi-scale mask classification domain adaptive network ［J］. Automotive Engineering， 2022， 44（9）： 1327-1338.
[17]	WANG Yan， CHEN Xiangyu， YOU Yurong， et al. Train in germany， test in the usa： making 3D object detectors generalize ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2020： 11713-11723.
[18]	YANG Jihan， SHI Shaoshuai， WANG Zhe， et al. St3D： self-training for unsupervised domain adaptation on 3d object detection ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2021： 10368-10378.
[19]	YANG Jihan， SHI Shaoshuai， WANG Zhe， et al. St3D++： denoised self-training for unsupervised domain adaptation on 3D object detection ［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2022， 45（5）： 6354-6371.
[20]	CHEN Zhuoxiao， LUO Yadan， WANG Zheng， et al. Revisiting domain-adaptive 3D object detection by reliable， diverse and class-balanced pseudo-labeling ［C］. Proceedings of the IEEE/CVF International Conference on Computer Vision， 2023： 3714-3726.
[21]	NEYSHABUR Behnam， SEDGHI Hanie， ZHANG Chiyuan. What is being transferred in transfer learning？［J］. Advances in Neural Information Processing Systems， 2020， 33： 512-523.
[22]	TSAI D， BERRIO J S， SHAN M， et al. MS3D++： ensemble of experts for multi-source unsupervised domain adaptation in 3D object detection ［J］. IEEE Transactions on Intelligent Vehicles， 2024： 1-16.

方法	AP_0.25/%	AP_0.5/%	AP_0.75/%
PointPillars式	90.12	89.96	89.13
PP-PBD	92.41	92.39	91.06
RoadPillars	93.15	92.45	91.72

方法	位置误差/m	角度误差/（°）
PointPillars式	0.165	1.56
PP-PBD	0.079	1.32
RoadPillars	0.094	1.12

域适应方法	迁移任务1： V2X-Seq to Dair-V2X （车辆目标 AP_0.5 BEV/3D）
域适应方法	Easy	Closed Gap	Mod.	Closed Gap	Hard	Closed Gap
Source only	78.20 / 29.52	-41.8%↓	64.52 / 25.61	-46.9%↓	64.50 / 25.72	-46.9%↓
SN^［17］	78.90 / 30.63	+3.4%↑	66.53 / 27.02	+6.1%↑	66.42 / 27.04	+5.8%↑
ST3D^［18］	79.08 / 29.25	+2.9%↑	66.91 / 26.02	+6.2%↑	66.76 / 25.96	+5.8%↑
ST3D++^［19］	79.76 / 31.88	+7.4%↑	67.29 / 27.82	+8.7%↑	67.21 / 27.88	+8.5%↑
ReDB^［20］	79.80 / 31.75	+7.4%↑	67.51 / 28.01	+9.4%↑	67.40 / 28.05	+9.1%↑
MS3D++^［22］	79.04 / 32.96	+5.7%↑	65.87 / 29.34	+6.5%↑	65.81 / 29.34	+6.3%↑
Ours （w/o SAC）	79.01 / 33.72	+6.2%↑	66.29 / 28.08	+6.5%↑	66.80 / 28.06	+7.7%↑
Ours （w SAC）	79.52 / 35.59	+9.5%↑	67.92 / 28.45	+10.8%↑	67.88 / 29.23	+11.3%↑
Oracle	92.44 / 92.44		84.92 / 84.92		84.92 / 84.92
域适应方法	迁移任务2： RCooper to Dair-V2X （车辆目标AP_0.5 BEV/3D）
域适应方法	Easy	Closed Gap	Mod.	Closed Gap	Hard	Closed Gap
Source only	24.69 / 23.62	-75.5%↓	24.72 / 24.67	-72.7%↓	24.72 / 24.67	-72.6%↓
SN^［17］	25.80 / 24.95	+1.7%↑	25.83 / 24.97	+1.1%↑	25.83 / 24.97	+1.1%↑
ST3D^［18］	28.55 / 28.52	+5.9%↑	28.58 / 28.55	+6.0%↑	28.58 / 28.55	+6.0%↑
ST3D++^［19］	29.47 / 29.44	+7.2%↑	29.50 / 29.47	+7.3%↑	29.50 / 29.47	+7.3%↑
ReDB^［20］	30.39 / 30.33	+8.4%↑	30.43 / 30.38	+8.7%↑	30.43 / 30.38	+8.7%↑
MS3D++^［22］	28.08 / 27.02	+4.6%↑	28.09 / 27.03	+4.4%↑	28.09 / 27.03	+4.4%↑
Ours （w/o SAC）	30.54 / 30.50	+8.6%↑	30.53 / 30.50	+8.9%↑	30.52 / 30.49	+8.9%↑
Ours （w SAC）	30.87 / 30.83	+9.0%↑	30.86 / 30.81	+9.4%↑	30.86 / 30.80	+9.4%↑

方法	ST3D++	ReDB	本方法
单帧推理耗时/ms	70.37	72.43	69.14
单轮离线训练/s	1 337	1 398	502
伪标签生成耗时/s	118	1 822	122