基于视觉的轻量化路面异常检测算法

doi:10.13229/j.cnki.jdxbgxb.20240802

摘要/Abstract

摘要：

为实现针对复杂路面多种异常状况的快速、准确检测，本文提出了一种行车视角下的路面异常检测算法ATFL-YOLOv8。首先，采用ADown卷积模块替换基线模型（YOLOv8n）中的部分普通卷积，进行高效的特征提取及采样；其次，在基线模型主干网络末端添加Triplet注意力层，提升模型的感知能力；然后，采用部分卷积思想构造出新的轻量化模块C2f-Faster，替换基线模型颈部网络中的C2f模块；再次，引入全新的LSCD-Head检测头进一步减少模型参数量并提升模型检测性能。测试结果显示：相较于基线模型，ATFL-YOLOv8在自建的行车视角下包含坑洞、裂缝、井盖、减速带的路面异常数据集上mAP0.5与mAP@0.5：0.95分别提升3.1个百分点、4.2个百分点，达到89.9%、59.7%；同时参数量、浮点运算数量、模型大小分别下降47%、37%、45%，降至1.61 M、5.2 G、3.31 MB，经实车验证，其具备一定低速下行车端路面异常检测能力。

关键词: 路面异常, 目标检测, 轻量化, 车辆工程

Abstract:

To achieve rapid and accurate detection of various abnormal conditions on complex road surfaces， a road surface anomaly detection algorithm ATFL-YOLOv8 is proposed from a driving perspective. Firstly， the ADown convolution module is used to replace some of the ordinary convolutions in the baseline model （YOLOv8n） for efficient feature extraction and sampling； secondly， a Triplet attention layer is added at the end of the baseline model backbone network to enhance the model's perceptual ability； again， using the idea of partial convolution， a new lightweight module C2f Faster is constructed to replace the C2f module in the neck network of the baseline model； finally， the introduction of a brand new LSCD Head detection head further reduces the number of model parameters and improves model detection performance. The test results show that compared to the baseline model， ATFL-YOLOv8 has mAP0.5 and mAP@0.5 0.95 increased by 3.1% and 4.2% respectively， reaching 89.9% and 59.7%； at the same time， the number of parameters， floating-point operations， and model size decreased by 47%， 37%， and 45%， respectively， to 1.61 M， 5.2 G， and 3.31 MB. Through actual vehicle verification， it has been proven that it has the ability to detect road surface abnormalities at certain low speeds.

Key words: abnormal road surface, object detection, lightweight, vehicle engineering

中图分类号:

U418.3

李军,杨飞帆,龚胜,周科宇. 基于视觉的轻量化路面异常检测算法[J]. 吉林大学学报(工学版), 2026, 56(2): 488-496.

Jun LI,Fei-fan YANG,Sheng GONG,Ke-yu ZHOU. Lightweight pavement anomaly detection algorithm based on vision[J]. Journal of Jilin University(Engineering and Technology Edition), 2026, 56(2): 488-496.

图/表 10

图1

图2

图3

图4

图5

表1

表2

表3

图6

图7

参考文献 16

[1]	交通运输部. 2022年交通运输行业发展统计公报[N]. 中国交通报, 2023-06-16(002).
[2]	Li P, Li H. Research on fod detection for airport runway based on YOLOv3[C]∥39th Chinese Control Conference (CCC), Shenyang,China, 2020: 7096-7099.
[3]	Shah S, Deshmukh C. Pothole and bump detection using convolution neural networks[C]∥ IEEE Transportation Electrification Conference (ITEC-India), Bangalore, India, 2019: 1-4.
[4]	Song W, Jia G, Zhu H, et al. Automated pavement crack damage detection using deep multiscale convolutional features[J]. Journal of Advanced Transportation, 2020, 2020(1): 6412562.
[5]	Kalfarisi R, Wu Z Y, Soh K. Crack detection and segmentation using deep learning with 3D reality mesh model for quantitative assessment and integrated visualization[J]. Journal of Computing in Civil Engineering, 2020, 34(3): 04020010.
[6]	Arya D, Maeda H, Ghosh S K, et al. RDD2020: an annotated image dataset for automatic road damage detection using deep learning[J]. Data in Brief, 2021, 36: 107133.
[7]	Arya D, Maeda H, Ghosh S K, et al. Rdd2022: a multi-national image dataset for automatic road damage detection[J/OL].[2024-06-25]. arXiv Preprint arXiv:.
[8]	Wang C Y, Yeh I H, Liao H Y M. YOLOv9: learning what you want to learn using programmable gradient information[J/OL].[2024-06-26].arXiv Preprint arXiv : 2402.13616.
[9]	Misra D, Nalamada T, Arasanipalai A U, et al. Rotate to attend: convolutional triplet attention module[C]∥Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, Hawaii, USA, 2021: 3139-3148.
[10]	Chen J, Kao S, He H, et al. Run, don't walk: chasing higher FLOPS for faster neural networks[C]∥ Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Hawaii, USA, 2023: 12021-12031.
[11]	Zeng J, Zhong H. YOLOv8-PD: an improved road damage detection algorithm based on YOLOv8n model[J]. Scientific Reports, 2024, 14(1): 12052.
[12]	Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]∥ Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, New York, USA, 2016: 779-788.
[13]	Lin T Y, Dollár P, Girshick R, et al. Feature pyramid networks for object detection[C]∥ Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, California, USA, 2017: 2117-2125.
[14]	Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C]∥ Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Nevada, USA, 2018: 7132-7141.
[15]	Woo S, Park J, Lee J Y, et al. CBAM: convolutional block attention module[C]∥ Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 2018: 3-19.
[16]	陈仁祥, 胡超超, 胡小林, 等. 基于改进 YOLOv5 的驾驶员分心驾驶检测[J]. 吉林大学学报: 工学版, 2024, 54(4): 959-968.
	Chen Ren-xiang, Hu Chao-chao, Hu Xiao-lin,et al.Driver distracted driving detection based on improved YOLOv5[J]. Journal of Jilin University (Engineering and Technology Edition), 2024, 54(4): 959-968.

相关文章 15

[1]	高镇海,鲍明喜,赵睿,唐明弘,高菲. 基于目标锚点驱动的多模态轨迹预测方法[J]. 吉林大学学报(工学版), 2026, 56(1): 21-30.
[2]	张向文,王子豪. 电动汽车制动模式切换过程电液协调控制策略[J]. 吉林大学学报(工学版), 2026, 56(1): 31-43.
[3]	冯志刚,任梦媛,董冰,于明月. 基于多频带特征图和改进SqueezeNet的滚动轴承故障诊断[J]. 吉林大学学报(工学版), 2026, 56(1): 96-108.
[4]	孙天骏,杨惠喆,蔡荣贵,冯嘉仪,冉锐,刘斌. 面向纯电动汽车自适应巡航系统的人性化起停控制策略[J]. 吉林大学学报(工学版), 2025, 55(9): 2847-2857.
[5]	兰巍,周政,王冠宇,王伟,张苗苗. 基于机器学习的汽车设计智能拟合方法[J]. 吉林大学学报(工学版), 2025, 55(9): 2858-2863.
[6]	朱冰,孟鹏翔,刘斌,韩嘉懿,赵健,陈志成,宋东鉴,陶晓文. 基于交通环境信息的虚拟车道线拟合方法[J]. 吉林大学学报(工学版), 2025, 55(9): 2935-2945.
[7]	李寿涛,贾湘怡,朱军,郭洪艳,于丁力. 基于Level-K的智能驾驶汽车无信控交叉路口决策方法[J]. 吉林大学学报(工学版), 2025, 55(9): 3069-3078.
[8]	高金武,孙少龙,王舜尧,高炳钊. 基于电机转矩补偿的增程器转速波动抑制策略[J]. 吉林大学学报(工学版), 2025, 55(8): 2475-2486.
[9]	刘长钊,宋健,李峥琪,张铁,王磊. 考虑动力学性能的高速薄壁齿轮多目标优化[J]. 吉林大学学报(工学版), 2025, 55(8): 2487-2500.
[10]	于贵申,陈鑫,唐悦,赵春晖,牛艾佳,柴辉,那景新. 激光表面处理对铝-铝粘接接头剪切强度的影响[J]. 吉林大学学报(工学版), 2025, 55(8): 2555-2569.
[11]	郭志荣,李刚. 基于高斯核密度估计的高速运动目标检测算法[J]. 吉林大学学报(工学版), 2025, 55(8): 2741-2745.
[12]	赵俊武,曲婷,胡云峰. 基于自适应采样的智能车辆轨迹规划方法[J]. 吉林大学学报(工学版), 2025, 55(8): 2802-2816.
[13]	宋学伟,于泽平,肖阳,王德平,袁泉,李欣卓,郑迦文. 锂离子电池老化后性能变化研究进展[J]. 吉林大学学报(工学版), 2025, 55(6): 1817-1833.
[14]	贾美霞,胡建军,肖凤. 基于多软件联合的车用电机变工况多物理场仿真方法[J]. 吉林大学学报(工学版), 2025, 55(6): 1862-1872.
[15]	肖纯,易子淳,周炳寅,张少睿. 基于改进鸽群优化算法的燃料电池汽车模糊能量管理策略[J]. 吉林大学学报(工学版), 2025, 55(6): 1873-1882.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

模型	mAP@0.5/%	mAP@0.5：0.95/%	Params/M	FLOPs/G	模型大小/MB
YOLOv8n	86.8	55.5	3.01	8.2	5.99
YOLOv8n-ShuffleNetV2	79.7	44.1	1.83	5.1	3.74
YOLOv8n-MoblieNetV3	83.5	48.0	2.35	5.8	4.79
YOLOv8n-Ghost	85.1	49.0	1.72	5.1	3.59
YOLOv8-ADown	88.1	55.7	2.60	7.5	5.21
YOLOv8-C2f-Faster	86.2	54.5	2.67	7.5	5.32
YOLOv8-LSCD-Head	88.9	57.8	2.37	6.6	4.74
YOLOv8-ADown-C2f-Faster-LSCD-Head	88.7	57.5	1.61	5.2	3.30

ADown	Treplet Attention	C2f-Faster	LSCD-Head	mAP@0.5/%	mAP@0.5：0.95/%	Params/M	FLOPs/G	模型大小/MB
				86.8	55.5	3.01	8.2	5.99
√				88.1	55.7	2.60	7.5	5.21
	√			89.1	55.5	3.01	8.2	6.00
		√		86.2	54.5	2.67	7.5	5.32
			√	88.9	57.8	2.37	6.6	4.74
√	√			89.9	56.5	2.60	7.5	5.21
√			√	89.6	57.6	1.95	5.9	3.96
	√	√		87.5	53.6	2.67	7.5	5.34
	√		√	89.3	59.1	2.37	6.6	4.75
√	√		√	90.5	59.6	1.95	5.9	3.97
√		√	√	88.7	57.5	1.61	5.2	3.30
√	√	√	√	89.9	59.7	1.61	5.2	3.31

模型	mAP@0.5/%	mAP@0.5：0.95/%	Params /M	FLOPs /G	模型大小/MB
YOLOv3-tiny	88.9	51.7	8.68	13.0	16.64
YOLOv5n	88.0	58.6	1.77	4.2	3.73
YOLOv6	83.8	47.9	4.24	11.9	8.32
YOLOv7-tiny	82.7	47.4	6.02	13.2	11.72
YOLOv8n	86.8	55.5	3.01	8.2	5.99
YOLOv8s	90.0	58.5	11.14	28.7	21.49
ATFL-YOLOv8	89.9	59.7	1.61	5.2	3.31