基于3DSSD的差异路口自适应联邦学习算法

doi:10.3969/j.issn.1674-8484.2024.05.011

摘要/Abstract

摘要：

在智能交通中，为弥补路侧端点云数据集的缺乏，提高目标检测模型泛化能力，提出了一种基于参数自适应联邦学习（FL）的点云目标检测算法（FLA3DSSD）。在各个路侧客户端数据不互通的情况下，将基于点的3D单级目标检测器（3DSSD）算法与经典联邦学习（FL）策略相结合，同时通过上传局部模型在路侧服务器进行模型自适应参数融合改进客户端模型参数更新策略，实现数据信息共享，并提升检测精度。结果表明：在车路协同差异路口场景算法部署任务中，对比本地数据训练模型，直接部署于其余客户端，用FL与3DSSD的聚合模型所测试的检测平均精度（AP），上升5%~40%；带有改进参数的自适应联邦学习FLA3DSSD聚合模型，实现AP值提升1%~7%。因此，在保护数据隐私的前提下，本方法能够提高模型泛化能力和检测精度。

关键词: 智能交通, 车路协同, 目标检测, 激光点云, 联邦学习（FL）策略, 深度学习模型

Abstract:

A point-cloud object-detection algorithm (FLA3DSSD) was proposed based on a parameter adaptive Federated Learning (FL) strategy to solve the problems of lack of roadside endpoint cloud dataset and the low generalization ability of object detection models. The point-cloud based 3-D Single-Stage ObjectDetector algorithm (3DSSD) was combined with the FL in the cases in which the data from various roadside clients are not interconnected. The client model parameter update strategy was improved by uploading local models to the server for model adaptive parameter fusion, to achieve data information sharing. The results show that the aggregation model combining classical federated learning and 3DSSD algorithm showed a 5%~40% increase in detection Average Precision (AP) compared to the locally trained model deployed directly to other clients for testing in the deployment task of the vehicle road collaborative differential intersection scene algorithm; Improved parameter adaptive federated learning FLA3DSSD achieves a 1%~7% increase in AP value based on the aggregated model. Therefore, the method improves the generalization ability and detection accuracy with protecting data privacy.

Key words: intelligent transportation, vehicle-road coordination, object detection, laser point-cloud, Federated-Learning (FL) strategy, Deep-Learning model

中图分类号:

TP183

石丽英, 周国峰, 李泽星, 曹莉凌. 基于3DSSD的差异路口自适应联邦学习算法[J]. 汽车安全与节能学报, 2024, 15(5): 732-741.

SHI Liying, ZHOU Guofeng, LI Zexing, CAO Liling. Adaptive federated learning algorithm for differential intersection based on 3DSSD[J]. Journal of Automotive Safety and Energy, 2024, 15(5): 732-741.

图/表 13

参考文献 14

[1]	WEN Jieluo, Urtasun R. Pixor: Real-time 3D object detection from point clouds[C]// Proc IEEE Conf Computer Vision Pattern Recog, Salt Lake City, UT, USA, 2018: 7652-7660.
[2]	YAN Yan, MAO Yuxing, LI Bo. Second: Sparsely embedded convolutional detection[J]. Sensors, 2018, 18(10): 3337.
[3]	Lang A H, Vora S, Caesar H, et al. Pointpillars: Fast encoders for object detection from point clouds[C]// Proc IEEE Conf Comput Visi Patt Recog, Long Beach, CA, USA, 2019: 12697-12705.
[4]	YANG Zetong, SUN Yanan, LIU Shu, et al. 3dssd: Point-based 3d single stage object detector[C]// Proc IEEE Conf Comput Visi Patt Recog, Seattle, WA, USA, 2020: 11040-11048.
[5]	McMahan B, Moore E, Ramage D, et al. Communication-efficient learning of deep networks from decentralized data[C]// Arti Intell Stat, AISTATS, Ft. Lauderdale, FL, USA, 2017: 1273-1282.
[6]	LIU Yang, HUANG Anbu, LUO Yu, et al. Fedvision:An online visual object detection platform powered by federated learning[C]// Proc AAAI Conf Arti Intell, New York, USA, 2020, 34(8): 13172-13179.
[7]	王生生, 路淑贞, 曹斌. 面向隐私保护联邦学习的医学影像目标检测算法[J]. 计算机辅助设计与图形学学报, 2021, 33(10): 1553-1562.
	WANG Shengsheng, LU Shuzhen, CAO Bin. Medical image object detection algorithm for privacy protection federated learning[J]. J Comput-Aid Des Comput Graph, 2021, 33(10): 1553-1562. (in Chinese)
[8]	赵羽, 杨洁, 刘淼, 等. 面向视频监控基于联邦学习的智能边缘计算技术[J]. 通信学报, 2020, 41(10): 109-115. doi: 10.11959/j.issn.1000-436x.2020192
	ZHAO Yu, YANG Jie, LIU Miao, et al. Intelligent edge computing technology based on federated l earning for video surveillance[J]. J Commu, 2020, 41(10): 109-115. (in Chinese)
[9]	ZHU Hangyu, XU Jinjin, LIU Shiqing, et al. Federated learning on non-IID data: A survey[J]. Neurocomputing, 2021, 465: 371-390.
[10]	LI Tian, Sahu A K, Zaheer M, et al. Federated optimization in heterogeneous networks[J]. Proc Mach Learn Syst, 2020, 2: 429-450.
[11]	HUANG Yutao, CHU Liyang, ZHOU Zirui, et al. Personalized cross-silo federated learning on non-iid data[C]// Proc AAAI Conf Arti Intell, New York, United States, 2021, 35(9): 7865-7873.
[12]	QI C. Ruizhongtai, LI Yi, SU Hao, et al. Pointnet++: Deep hierarchical feature learning on point sets in a metric space[C/OL]// Annu Conf Neur Info Proc Syst, NIPS, 2017, 30. (2017-01-01).https://proceedings.neurips.cc/paper/2017/hash/d8bf84be3800d12f74d8b05e9b89836f-Abstract.html.
[13]	TIAN Zhi, SHEN Chunhua, CHEN Hao, et al. FCOS: A simple and strong anchor-free object detector[J]. IEEE Trans Patt Anal Mach Intel, 2020, 44(4): 1922-1933.
[14]	YU Haibao, YANG Wenxian, RUAN Hongzhi, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting[C]// Proc IEEE Conf Compu Visi Patt Recog, Vancouver, Canada, 2023: 5486-5495.

数据集编号	数据量/帧	场景特征
1	1 945	五车道(直行)
2	2 730	五车道变四车道(直行、转弯)
3	2 899	三车道变双车道(直行、转弯，右转单独)
4	1 850	四车道变三车道(直行、转弯)
5	1 035	五车道变四车道(直行、转弯，右转车道可直行)
6	816	单侧单车道单向(右转单独，有安全岛)

数据集编号	数据量/帧	场景特征
1	1 945	五车道(直行)
2	2 730	五车道变四车道(直行、转弯)
3	2 899	三车道变双车道(直行、转弯，右转单独)
4	1 850	四车道变三车道(直行、转弯)
5	1 035	五车道变四车道(直行、转弯，右转车道可直行)
6	816	单侧单车道单向(右转单独，有安全岛)

数据集编号	AP_3D/%		AP_BEV/%
数据集编号	IoU = (0.7，0.7，0.7)	IoU = (0.7，0.5，0.5)	IoU = (0.7，0.7，0.7)	IoU = (0.7，0.5，0.5)
1	88.24，87.77，80.48	90.81，90.71，90.67	90.05，89.50，89.43	99.40，90.80，90.77
2	78.89，62.78，62.78	81.39，72.29，72.29	81.02，71.60，71.60	81.65，72.54，72.54
3	64.57，41.17，41.24	70.15，52.78，52.74	69.39，52.27，52.13	70.39，52.94，52.90
4	53.70，34.85，34.85	54.25，36.13，36.12	54.10，35.78，35.78	54.25，36.13，36.12
5	42.05，36.10，35.97	80.82，72.19，72.19	58.12，54.66，54.65	81.22，72.44，72.41
6	40.49，33.29，39.99	62.03，44.62，53.46	52.76，42.88，44.42	63.74，44.39，49.50

数据集编号	AP_3D/%		AP_BEV/%
数据集编号	IoU = (0.7，0.7，0.7)	IoU = (0.7，0.5，0.5)	IoU = (0.7，0.7，0.7)	IoU = (0.7，0.5，0.5)
1	88.24，87.77，80.48	90.81，90.71，90.67	90.05，89.50，89.43	99.40，90.80，90.77
2	78.89，62.78，62.78	81.39，72.29，72.29	81.02，71.60，71.60	81.65，72.54，72.54
3	64.57，41.17，41.24	70.15，52.78，52.74	69.39，52.27，52.13	70.39，52.94，52.90
4	53.70，34.85，34.85	54.25，36.13，36.12	54.10，35.78，35.78	54.25，36.13，36.12
5	42.05，36.10，35.97	80.82，72.19，72.19	58.12，54.66，54.65	81.22，72.44，72.41
6	40.49，33.29，39.99	62.03，44.62，53.46	52.76，42.88，44.42	63.74，44.39，49.50

测试集编号	AP_3D/%		AP_BEV/%
测试集编号	IoU = (0.7，0.7，0.7)	IoU = (0.7，0.5，0.5)	IoU = (0.7，0.7，0.7)	IoU = (0.7，0.5，0.5)
1	78.06，77.83，76.04	90.56，90.29，90.14	88.91，87.14，80.30	90.69，90.62，90.55
2	72.21，63.18，63.18	81.33，72.24，72.24	80.14，71.06，71.06	81.56，72.46，72.46
3	21.69，20.25，20.20	57.46，41.71，41.66	46.37，38.20，38.06	67.21，50.68，50.53
4	51.10，27.04，27.04	54.41，35.98，35.98	54.03，33.56，33.56	54.41，36.01，36.01
5	50.78，42.35，42.34	72.19，63.16，63.11	62.20，59.98，53.47	72.34，63.31，63.26
6	23.04，18.52，20.30	44.33，32.32，39.28	31.49，23.38，29.64	46.04，32.85，40.14