A lightweight rotation-invariant network for LiDAR-based place recognition

doi:10.11947/j.AGCS.2025.20230302

Abstract

Abstract:

Accurate place recognition using LiDAR is critical for achieving global localization in domains such as autonomous driving or robot navigation. While the state-of-the-art methods focus on enhancing the feature encoding capability, they often neglect the essential requirement of maintaining rotation invariance in the place recognition process. Additionally, these methods suffer from challenges, such as the large model size and their dependence on complex preprocessing procedures. To address these challenges, this paper introduces RIP-Net, a super lightweight neural network designed for rotation-invariant place recognition of point clouds. Firstly, RIP-Net gathers point clusters of local regions and constructs basic rotation-invariant features; Secondly, the residual structures and attention mechanism are employed to enhance the perception of local regions by incorporating multi-scale information; Finally, we utilized the generalized-mean pooling function to aggregate global feature, and place recognition is achieved based on the feature distance. The experimental results on 4 large-scale point cloud datasets demonstrate that RIP-Net not only achieves rotation invariance but also outperforms existing methods in terms of various accuracy metrics. Moreover, the parameter number of the IR-Net is 0.3 million, which is significantly lower compared to existing methods. Experimental results also demonstrate that RIP-Net can achieve accurate place recognition directly using large-scale raw point clouds without any data preprocessing steps. These findings underscore the practical value and promising applications of the proposed RIP-Net method.

Key words: LiDAR-based place recognition, light weight network, rotation-invariant, global localization, deep learning

CLC Number:

P237

Zhenghua ZHANG, Guoliang CHEN. A lightweight rotation-invariant network for LiDAR-based place recognition[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(1): 90-103.

Figures/Tables 20

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Fig. 6

Tab. 1

Tab. 2

Tab. 3

Fig. 7

Fig. 8

Fig. 9

Tab. 4

Fig. 10

Fig. 11

Fig. 12

Tab. 5

Fig. 13

Fig. 14

Fig. 15

References 30

[1]	戴德云. 基于多元信息融合的场景多层级识别方法研究[D]. 合肥: 中国科学技术大学, 2022.
	DAI Deyun. Research on multi-level scene recognition method based on multi-information fusion[D]. Hefei: University of Science and Technology of China, 2022.
[2]	杨元喜. 弹性PNT基本框架[J]. 测绘学报, 2018, 47(7): 893-898. DOI:. doi: 10.11947/J.AGCS.2018.20180149
	YANG Yuanxi. Resilient PNT concept frame[J]. Acta Geodaetica et Cartographica Sinica, 2018, 47(7): 893-898. DOI:. doi: 10.11947/j.AGCS.2018.20180149
[3]	张星, 林静, 李清泉, 等. 结合感知哈希与空间约束的室内连续视觉定位方法[J]. 测绘学报, 2021, 50(12): 1639-1649.DOI:. doi: 10.11947/J.AGCS.2021.20200286
	ZHANG Xing, LIN Jing, LI Qingquan, et al. Indoor continuous visual positioning method combining perceptual hashing and spatial constraints[J]. Acta Geodaetica et Cartographica Sinica, 2021, 50(12): 1639-1649.DOI:. doi: 10.11947/J.AGCS.2021.20200286
[4]	张恒才, 蔚保国, 秘金钟, 等. 综合PNT场景增强系统研究进展及发展趋势[J]. 武汉大学学报(信息科学版), 2023, 48(4): 491-505.
	ZHANG Hengcai, YU Baoguo, BEI Jinzhong, et al. A survey of scene-based augmentation systems for conprehensive PNT[J]. Geomatics and Information Science of Wuhan University, 2023, 48(4): 491-505.
[5]	张正华. 基于深度学习的激光雷达点云实时配准与场景识别算法研究[D]. 徐州: 中国矿业大学, 2022.
	ZHANG Zhenghua. Study on deep-learning-based real-time registration and place recognition algorithm for LiDAR point cloud[D]. Xuzhou: China University of Mining and Technology, 2022.
[6]	WOHLKINGER W, VINCZE M. Ensemble of shape functions for 3D object classification[C]//Proceedings of 2011 IEEE International Conference on Robotics and Biomimetics. Karon Beach: IEEE, 2011.
[7]	HE Li, WANG Xiaolong, ZHANG Hong. M2DP: a novel 3D point cloud descriptor and its application in loop closure detection[C]//Proceedings of 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems. Daejeon: IEEE, 2016: 231-237.
[8]	KIM G, KIM A. Scan context: egocentric spatial descriptor for place recognition within 3D point cloud map[C]//Proceedings of 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems. Madrid: IEEE, 2018: 4802-4809.
[9]	杨必胜, 董震. 点云智能研究进展与趋势[J]. 测绘学报, 2019, 48(12): 1575-1585.DOI:. doi: 10.11947/j.AGCS.2019.20190465
	YANG Bisheng, DONG Zhen. Progress and perspective of point cloud intelligence[J]. Acta Geodaetica et Cartographica Sinica, 2019, 48(12): 1575-1585.DOI:. doi: 10.11947/j.AGCS.2019.20190465
[10]	UY M A, LEE G H. PointNetVLAD: deep point cloud based retrieval for large-scale place recognition[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018.
[11]	CHARLES R Q, HAO Su, MO Kaichun, et al. PointNet: deep learning on point sets for 3D classification and segmentation[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017.
[12]	ARANDJELOVIC R, GRONAT P, TORII A, et al. NetVLAD: CNN architecture for weakly supervised place recognition[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016.
[13]	ZHANG Wenxiao, XIAO Chunxia. PCAN: 3D attention map learning using contextual information for point cloud based retrieval[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019.
[14]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[EB/OL]. [2023-12-10]. https://doi.org/10.48550/arXiv.1706.03762.
[15]	LIU Zhe, ZHOU Shunbo, SUO Chuanzhe, et al. LPD-net: 3D point cloud learning for large-scale place recognition and environment analysis[C]//Proceedings of 2019 IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019.
[16]	HUI L, CHENG M, XIE J, et al. Efficient 3D point cloud feature learning for large-scale place recognition[J]. IEEE Transactions on Image Processing, 2022, 31: 1258-1270.
[17]	KOMOROWSKI J. MinkLoc3D: point cloud based large-scale place recognition[C]//Proceedings of 2021 IEEE Winter Conference on Applications of Computer Vision. Waikoloa: IEEE, 2021.
[18]	RADENOVIC F, TOLIAS G, CHUM O. Fine-tuning CNN image retrieval with no human annotation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(7): 1655-1668.
[19]	XU Tianxing, GUO Yuanchen, LI Zhiqiang, et al. TransLoc3D: point cloud based large-scale place recognition using adaptive receptive fields[J]. Communications in Information and Systems, 2023, 23(1): 57-83.
[20]	FAN Z, SONG Z, LIU H, et al. SVT-Net: super light-weight sparse voxel transformer for large scale place recognition[J]. AAAI Conference on Artificial Intelligence. 2022, 36(1), 551-560.
[21]	TIAN Gengxuan, ZHAO Junqiao, CAI Yingfeng, et al. VNI-Net: vector neurons-based rotation-invariant descriptor for LiDAR place recognition[EB/OL]. [2023-10-15]. https://arxiv.org/abs/2308.12870v1.
[22]	FAN Zhaoxin, SONG Zhenbo, LIU Hongyan, et al. SVT-net: super light-weight sparse voxel transformer for large scale place recognition[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2022, 36(1): 551-560.
[23]	XIA Yan, XU Yusheng, LI Shuang, et al. SOE-net: a self-attention and orientation encoding network for point cloud based place recognition[C]//Proceedings of 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Nashville: IEEE, 2021.
[24]	MADDERN W, PASCOE G, LINEGAR C, et al. 1 year, 1000 km: the Oxford RobotCar dataset[J]. The International Journal of Robotics Research, 2017, 36(1): 3-15.
[25]	KIM G, PARK Y S, CHO Y, et al. MulRan: multimodal range dataset for urban place recognition[C]//Proceedings of 2020 IEEE International Conference on Robotics and Automation. Paris: IEEE, 2020.
[26]	LAZANYI K. Are we ready for self-driving cars-a case of principal-agent theory[C]//Proceedings of 2018 IEEE International Symposium on Applied Computational Intelligence and Informatics. Timisoara: IEEE, 2018.
[27]	HUI Le, YANG Hang, CHENG Mingmei, et al. Pyramid point cloud transformer for large-scale place recognition[C]//Proceedings of 2021 IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021.
[28]	ZHANG Z, CHEN G, SHU M, et al. LWR-Net: robust and lightweight place recognition network for noisy and low-density point clouds[J]. Sensors (Basel, Switzerland), 2023, 23(21): 8664.
[29]	KOMOROWSKI J, WYSOCZANSKA M, TRZCINSKI T. MinkLoc++: LiDAR and monocular image fusion for place recognition[C]//Proceedings of 2021 International Joint Conference on Neural Networks. Shenzhen: IEEE, 2021.
[30]	ZHANG W, QI J, WAN P, et al. An easy-to-use airborne LiDAR data filtering method based on cloth simulation[J]. Remote Sensing. 2016, 8(6), 501.

数据集名称	构建时间	构建机构	场景特征	训练集场景数	测试集场景数
Oxford RobotCar	2017	牛津大学	牛津市中心区域场景	23 177	3030
NUS Inhouse	2018	新加坡国立大学	大学校园、居民区与商业区场景	0	10 813
Mul Ran	2020	韩国科学技术院	韩国大田会议中心、科学院及河堤场景	26 799	43 749
KITTI odometry	2013	德国卡尔斯鲁厄理工学院	城市、乡村与公路等场景	0	18 266

方法	参数量	运算时间/ms	准确率/（%）	最近邻召回率/（%）	前1%召回率/（%）	F₁值
PointNetVLAD	19.78×10⁶	15	11.96	15.01	31.31	0.20
PCAN	20.42×10⁶	55	12.82	14.81	32.81	0.21
EPC-Net	4.70×10⁶	26	59.97	59.78	80.93	0.74
MinkLoc3D	1.06×10⁶	12	85.62	74.26	90.14	0.91
MinkLoc++	1.06×10⁶	12	73.70	80.79	92.75	0.84
PPT-Net	13.39×10⁶	22	61.03	61.55	82.82	0.75
SVT-Net	0.94×10⁶	11	77.23	72.21	89.41	0.86
LWR-Net	0.44×10⁶	10	41.39	40.93	62.80	0.57
SOE-Net	19.40×10⁶	22	86.56	86.72	95.60	0.92
RPR-Net	1.10×10⁶	238	—	81.00	92.20	—
VNI-Net	2.20×10⁶	574	—	85.50	94.40	—
RIP-Net	0.30×10⁶	9	87.50	87.57	95.83	0.93

方法	大学校园场景			居民区场景			商业区场景
方法	最近邻召回率/（%）	前1%召回率/（%）	F₁值	最近邻召回率/（%）	前1%召回率/（%）	F₁值	最近邻召回率/（%）	前1%召回率/（%）	F₁值
PointNetVLAD	18.16	33.76	0.20	15.17	28.42	0.20	17.39	26.12	0.27
PCAN	10.56	25.00	0.18	11.95	23.31	0.17	12.99	19.42	0.22
EPC-Net	60.10	80.56	0.75	51.89	70.77	0.68	57.48	69.44	0.73
MinkLoc3D	57.17	68.27	0.81	44.88	60.54	0.77	46.13	63.79	0.82
MinkLoc++	65.91	81.83	0.74	65.73	75.94	0.63	57.11	76.20	0.63
PPT-Net	42.53	69.08	0.65	54.83	68.14	0.60	40.59	61.56	0.73
SVT-Net	53.47	69.00	0.79	52.28	68.34	0.76	67.44	78.76	0.81
LWR-Net	41.80	64.42	0.60	41.66	62.96	0.58	56.22	68.76	0.73
SOE-Net	77.32	90.71	0.87	76.08	89.78	0.86	76.35	81.32	0.89
VNI-Net	85.30	95.00	—	83.30	91.50	—	81.40	86.80	—
RPR-Net	83.20	94.50	—	83.30	91.30	—	80.40	86.40	—
RIP-Net	85.35	95.31	0.91	83.49	92.10	0.89	77.96	84.52	0.87

方法	韩国大田会议中心场景			河堤场景			韩国科学技术院场景
方法	准确率/（%）	召回率/（%）	F₁值	准确率/（%）	召回率/（%）	F₁值	准确率/（%）	召回率/（%）	F₁值
MinkLoc3D	10.37	35.81	0.17	8.95	31.72	0.14	9.12	21.56	0.14
MinkLoc++	12.54	33.24	0.19	9.86	31.43	0.15	10.55	38.21	0.17
SVT-Net	11.21	40.79	0.18	8.37	34.62	0.13	12.45	39.18	0.19
RIP-Net	92.17	99.31	0.96	88.37	99.19	0.94	90.66	98.79	0.95

场景序列	准确率/（%）	召回率/（%）	F₁值
00	66.02	98.50	0.79
02	66.13	98.48	0.79
05	62.54	98.17	0.77
06	69.83	99.63	0.82
07	72.49	99.77	0.84
07	72.49	99.77	0.84