基于迁移学习和多策略改进的山区公路落石目标检测

杨磊; 张浩

doi:10.19509/j.cnki.dzkq.tb202603033

基于迁移学习和多策略改进的山区公路落石目标检测

doi: 10.19509/j.cnki.dzkq.tb202603033

杨磊^1, ,,
张浩²

河南信息统计职业学院建筑工程学院，郑州 450008
郑州大学土木工程学院，郑州 450001

基金项目: 国家自然科学基金项目（51608490）

详细信息

通讯作者:
E-mail：tougao4321@yeah.net

中图分类号: TP391.41；P642.21
计量
- 文章访问数: 114
- PDF下载量: 9
- 被引次数: 0
出版历程
- 收稿日期: 2026-03-21
- 录用日期: 2026-04-15
- 修回日期: 2026-04-12
- 网络出版日期: 2026-05-12

Rockfall object detection on mountain roads based on transfer learning and multi-strategy improvements

YANG Lei^{1
, ,},
ZHANG Hao²

Department of Architectural Engineering, Henan Vocational College of Information and Statistics, Zhengzhou 450008, China
College of Civil Engineering, Zhengzhou University, Zhengzhou 450001, China

More Information

Corresponding author: E-mail：tougao4321@yeah.net

摘要

摘要:
针对山区公路落石检测中真实样本稀缺、落石尺度多变以及野外复杂背景干扰等难题，现有模型在小样本条件下泛化能力弱、漏检率高，难以满足智能监测需求。本研究旨在构建一种能够适应小样本、多尺度与复杂环境的落石目标检测模型，提升检测精度，为山区公路落石智能监测预警提供技术支撑。本研究提出一种融合迁移学习与多策略改进的YOLOv8落石检测模型。首先，构建包含3000余张落石图像的数据集，涵盖多种岩性、不同光照条件及变化背景，数据来源于实地监控采集、手机拍摄以及数据增强操作，从小样本角度扩大数据多样性。模型设计方面，一方面引入基于大规模数据集的迁移学习，缓解小样本导致的特征学习不充分问题；另一方面在特征网络中嵌入坐标注意力机制（CA），强化模型对落石空间位置的响应，抑制复杂背景干扰；同时采用双向特征金字塔网络（BiFPN）代替原有特征融合路径，改善多尺度特征传递与融合效果，提升对不同尺度落石的感知能力；最后以高效交并比（EIoU）损失函数优化边界框回归，同时考虑中心点距离、长宽比和重叠区域，提高定位精度。实验结果显示，与基准YOLOv8模型相比，本模型的精确率提高17.1%，召回率提高24.7%，mAP50提升17.4%，3项指标均有显著改善。与其他轻量级模型横向对比，本模型mAP50达到74.6%，明显优于YOLOv5n的51.3%、YOLOv6n的56.7%及YOLOv8n的57.2%，在小样本条件下显著降低了漏检率和误报率，且较好地平衡了精度与效率。在测试集上的验证表明，模型在多种复杂背景和不同岩性场景下均表现出良好的泛化能力和鲁棒性。通过迁移学习与多策略改进，本研究提出的YOLOv8模型有效克服了山区公路落石检测中的样本稀缺、尺度多变和背景复杂问题，检测精度和鲁棒性表现突出，为山区公路落石智能监测预警系统的开发提供了可行的技术方案。
- 落石检测 /
- 迁移学习 /
- YOLOv8 /
- 坐标注意力机制（CA） /
- 双向特征金字塔 /
- 小样本学习
Abstract:
Objective
To address the challenges of scarce real-world rockfall samples, highly variable target scales, and complex background interference in mountain road rockfall detection, existing models suffer from weak generalization ability and high missed detection rates under small-sample conditions. Consequently, they fail to meet intelligent monitoring requirements. This study aims to construct a rockfall object detection model capable of adapting to small samples, multiple scales, and complex environments, thereby improving detection accuracy and providing technical support for intelligent rockfall monitoring and early warning on mountain roads.
Methods
This study proposed a YOLOv8 rockfall detection model integrating transfer learning and multi-strategy improvements. First, a dataset containing over 3 000 rockfall images was constructed, covering various rock lithologies, different lighting conditions, and varying backgrounds. Data sources included on-site surveillance camera captures, mobile phone photography, and data augmentation operations, thereby expanding data diversity from a small-sample perspective. In terms of model design, transfer learning based on a large-scale dataset was introduced to alleviate insufficient feature learning caused by limited samples. Additionally, the coordinate attention mechanism was embedded into the feature network to enhance the model's response to the spatial positions of rockfalls while suppressing complex background interference. Meanwhile, a weighted bidirectional feature pyramid network (BiFPN) was adopted to replace the original feature fusion path, improving multi-scale feature flow and enhancing the perception of rockfalls at different scales. Finally, the EIoU loss function was employed to optimize bounding box regression, simultaneously considering center point distance, aspect ratio, and overlap area to improve localization accuracy.
Results
Experimental results showed that compared with the baseline YOLOv8 model, the precision of the proposed model increased by 17.1%, recall by 24.7%, and mAP50 by 17.4%, with significant improvements in all three indicators. In comparison with other lightweight models, the proposed model achieved an mAP50 of 74.6%, significantly outperforming YOLOv5n (51.3%), YOLOv6n (56.7%), and YOLOv8n (57.2%). Under small-sample conditions, the missed detection rate and false alarm rate were substantially reduced, and a good balance between accuracy and efficiency was maintained. Validation on the test set demonstrated that the model exhibited good generalization ability and robustness across various complex backgrounds and different rock lithology scenarios.
Conclusion
Through transfer learning and multi-strategy improvements, the proposed YOLOv8 model effectively overcomes the problems of sample scarcity, scale variability, and background complexity in mountain road rockfall detection, achieving outstanding detection accuracy and robustness. It provides a feasible technical solution for the development of intelligent rockfall monitoring and early warning systems for mountain roads.
- rockfall detection /
- transfer learning /
- YOLOv8 /
- coordinate attention mechanism /
- BiFPN /
- few-shot learning

HTML全文

图 1 训练集标注结果

图b中颜色越深表示所属的边界框数量越多；（x，y）为边界框归一化中心点坐标；w为边界框归一化宽度；h为边界框归一化高度；下同

Figure 1. Annotation results of training set

下载: 全尺寸图片幻灯片

图 2 改进模块框架图

C为特征图通道数；H为特征图高度；W为特征图宽度；r为通道压缩倍率；P₃～P₇为5个不同尺度特征层，对应输入图像下采样2³～2⁷倍后的特征图

Figure 2. Framework of improved modules

下载: 全尺寸图片幻灯片

图 3 迁移学习流程

CA. 坐标注意力机制；BiFPN. 双向特征金字塔网络；EloU. 高效交并比；SPPF. 快速空间金字塔池化；SiLU. Sigmoid线性激活函数；C2f模块. 双卷积流式跨阶段局部模块；L₁～L₁₀. 骨干网络层编号；下同

Figure 3. Transfer learning workflow

下载: 全尺寸图片幻灯片

图 4 训练过程中性能指标的变化

mAP50. IoU阈值取0.5时的mAP；mAP50:95. IoU阈值在0.5～0.95之间、步长 0.05 下的mAP；IoU. 交并比；mAP. 平均精度均值；下同

Figure 4. Variation of performance indicators during training

下载: 全尺寸图片幻灯片

图 5 模型性能评价

Figure 5. Model performance evaluation

下载: 全尺寸图片幻灯片

图 6 对于测试集的目标检测结果（luoshi 0.8等均为落石目标编号，下同）

Figure 6. Object detection results on test set

下载: 全尺寸图片幻灯片

图 7 检测效果对比（a，b. YOLOv8；c，d. 本研究模型）

Figure 7. Comparison of detection results

下载: 全尺寸图片幻灯片

图 8 小目标检测性能评估

Figure 8. Performance evaluation of small object detection

下载: 全尺寸图片幻灯片

表 1 落石数据集

Table 1. Rockfall dataset

划分维度	样本分布
岩性	花岗岩（1024张）	灰岩（530张）	页岩（745张）
岩性	砂岩（808张）	其他（240张）
大小	小目标（708张）	中目标（1783张）	大目标（856张）
光照条件	晴朗白昼（2071张）	降雨白昼落石（676张）	夜晚落石（600张）
背景	普通背景（1908张）	植被干扰背景落石（874张）	人车干扰背景落石（565张）

下载: 导出CSV

表 2 超参数设计

Table 2. Hyperparameter settings

超参数名称	数值	超参数名称	数值
学习率	0.01	迭代次数	170
图像分辨率	640×640	权重衰减	0.0005
批次大小	16	并行进程	24

下载: 导出CSV

表 3 迁移学习模型消融实验结果

Table 3. Ablation test results of transfer learning model

编号	模型配置	P/%	R/%	mAP50/%	参数量/M	GFLOPs
1	YOLOv8（基准，从头训练）	58.3	52.6	57.2	3.2	8.9
2	+迁移学习	63.5	58.4	62.8	3.2	8.9
3	+迁移学习+CA	67.8	62.5	66.4	3.45	9.5
4	+迁移学习+CA+BiFPN	71.2	66.8	70.1	3.85	10.8
5	+迁移学习+CA+BiFPN+EIoU	75.4	77.3	74.6	4.05	11.3
注：P为精确率；R为召回率；GFLOPs为模型浮点计算量；下同

下载: 导出CSV

表 4 本研究迁移学习模型与其他模型的比较

Table 4. Comparison between transfer learning model in this study and other models

模型	P/%	R/%	mAP50/%	参数量/M	GFLOPs
YOLOv5n	55.2	48.6	51.3	1.9	4.5
YOLOv6n	59.8	53.2	56.7	4.7	11.4
YOLOv8n	58.3	52.6	57.2	3.2	8.9
YOLOv9n	62.5	57.4	61.8	2.0	7.7
YOLOv10n	68.2	67.5	65.4	2.3	8.0
本研究模型	75.4	77.3	74.6	4.05	11.3

下载: 导出CSV

参考文献(40)

[1]	TAN D Y, YIN J H, ZHU Z H, et al. Fast door-opening method for quick release of rock boulder or debris in large-scale physical model[J]. International Journal of Geomechanics, 2020, 20(2): 06019019. doi: 10.1061/(ASCE)GM.1943-5622.0001556
[2]	许浒, 邹鹏, 余志祥, 等. 山区公路高陡边坡引导式柔性缓冲系统的设计方法[J]. 中国公路学报, 2022, 35(9): 235-246. doi: 10.3969/j.issn.1001-7372.2022.09.018 XU H, ZOU P, YU Z X, et al. Design approach of guided flexible protection system for high and steep slope of mountain highways[J]. China Journal of Highway and Transport, 2022, 35(9): 235-246. (in Chinese with English abstract) doi: 10.3969/j.issn.1001-7372.2022.09.018
[3]	张乐, 陈沛, 向波, 等. 基于运动学三维模拟的山区公路崩塌灾害损伤评价及治理[J]. 地质科技通报, 2025, 44(2): 104-115. doi: 10.19509/j.cnki.dzkq.tb20240068 ZHANG L, CHEN P, XIANG B, et al. Rockfall damage evaluation and treatment suggestions for mountain highways based on three-dimensional kinematic simulations[J]. Bulletin of Geological Science and Technology, 2025, 44(2): 104-115. (in Chinese with English abstract) doi: 10.19509/j.cnki.dzkq.tb20240068
[4]	卢彦丞, 李军, 梁风, 等. 岩溶典型区崩塌落石被动防护网失效概率模拟[J]. 地质科技通报, 2024, 43(3): 240-250. doi: 10.19509/j.cnki.dzkq.tb20230552 LU Y C, LI J, LIANG F, et al. Failure probability simulation of passive protection net for collapses and rockfalls in typical karst area[J]. Bulletin of Geological Science and Technology, 2024, 43(3): 240-250. (in Chinese with English abstract) doi: 10.19509/j.cnki.dzkq.tb20230552
[5]	LIENHART W. Geotechnical monitoring using total stations and laser scanners: Critical aspects and solutions[J]. Journal of Civil Structural Health Monitoring, 2017, 7(3): 315-324. doi: 10.1007/s13349-017-0228-5
[6]	RAGNOLI M, SCARSELLA M, LEONI A, et al. Wireless sensor network-based rockfall and landslide monitoring systems: A review[J]. Sensors, 2023, 23(16): 7278. doi: 10.3390/s23167278
[7]	JANERAS M, JARA J A, ROYÁN M J, et al. Multi-technique approach to rockfall monitoring in the Montserrat massif (Catalonia, NE Spain)[J]. Engineering Geology, 2017, 219: 4-20. doi: 10.1016/j.enggeo.2016.12.010
[8]	刘飞跃, 杨天鸿, 朱万成, 等. 基于落石视频监测的露天矿岩质边坡临滑预警研究[J]. 岩土工程学报, 2025, 47(4): 792-800. doi: 10.11779/CJGE20240026 LIU F Y, YANG T H, ZHU W C, et al. Early warning for landslide of rock slopes in open-pit mine based on rockfall video monitoring[J]. Chinese Journal of Geotechnical Engineering, 2025, 47(4): 792-800. (in Chinese with English abstract) doi: 10.11779/CJGE20240026
[9]	LIU L, OUYANG W L, WANG X G, et al. Deep learning for generic object detection: A survey[J]. International Journal of Computer Vision, 2020, 128(2): 261-318. doi: 10.1007/s11263-019-01247-4
[10]	REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149. doi: 10.1109/TPAMI.2016.2577031
[11]	ZAIDI S S A, ANSARI M S, ASLAM A, et al. A survey of modern deep learning based object detection models[J]. Digital Signal Processing, 2022, 126: 103514. doi: 10.1016/j.dsp.2022.103514
[12]	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Anon. 2014 IEEE Conference on Computer Vision and Pattern Recognition. Columbus, OH, USA: IEEE, 2014: 580-587.
[13]	GIRSHICK R. Fast R-CNN[C]//Anon. 2015 IEEE International Conference on Computer Vision (ICCV). Santiago, Chile: IEEE, 2015: 1440-1448.
[14]	SUN P, ZHANG R, JIANG Y, et al. Sparse R-CNN: End-to-End object detection with learnable proposals[C]//Anon. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. [S. l. ]: [s. n. ], 2021: 14454-14463.
[15]	REDMON J, DIVVALA S, GIRSHICK R, et al. You Only Look Once: Unified, real-time object detection[C]//Anon. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, USA: IEEE, 2016: 779-788.
[16]	SAPKOTA R, FLORES-CALERO M, QURESHI R, et al. YOLO advances to its genesis: A decadal and comprehensive review of the You Only Look Once (YOLO) series[J]. Artificial Intelligence Review, 2025, 58(9): 274. doi: 10.1007/s10462-025-11253-3
[17]	REDMON J, FARHADI A. YOLO9000: Better, faster, stronger[C]//Anon. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI, USA: IEEE, 2017: 6517-6525.
[18]	ALI M L, ZHANG Z. The YOLO framework: A comprehensive review of evolution, applications, and benchmarks in object detection[J]. Computers, 2024, 13(12): 336. doi: 10.3390/computers13120336
[19]	MOHAMMED S Y. Architecture review: Two-stage and one-stage object detection[J]. Franklin Open, 2025, 12: 100322. doi: 10.1016/j.fraope.2025.100322
[20]	DU Y C, PAN N, XU Z H, et al. Pavement distress detection and classification based on YOLO network[J]. International Journal of Pavement Engineering, 2021, 22(13): 1659-1672. doi: 10.1080/10298436.2020.1714047
[21]	LI J L, YUAN C L, WANG X F. Real-time instance-level detection of asphalt pavement distress combining space-to-depth (SPD) YOLO and omni-scale network (OSNet)[J]. Automation in Construction, 2023, 155: 105062. doi: 10.1016/j.autcon.2023.105062
[22]	GUO K Y, HE C B, YANG M, et al. A pavement distresses identification method optimized for YOLOv5s[J]. Scientific Reports, 2022, 12: 3542. doi: 10.1038/s41598-022-07527-3
[23]	FLORES-CALERO M, ASTUDILLO C A, GUEVARA D, et al. Traffic sign detection and recognition using YOLO object detection algorithm: A systematic review[J]. Mathematics, 2024, 12(2): 297. doi: 10.3390/math12020297
[24]	邢岩, 郭思豪, 张振, 等. 基于CGT-YOLO的小目标交通标志识别算法[J]. 华南理工大学学报(自然科学版), 2026, 54(3): 65-78. doi: 10.12141/j.issn.1000-565X.250092 XING Y, GUO S H, ZHANG Z, et al. CGT-YOLO-based algorithm for small-target traffic sign recognition[J]. Journal of South China University of Technology (Natural Science Edition), 2026, 54(3): 65-78. (in Chinese with English abstract) doi: 10.12141/j.issn.1000-565X.250092
[25]	WANG C, ZHENG B, LI C X. Efficient traffic sign recognition using YOLO for intelligent transport systems[J]. Scientific Reports, 2025, 15: 13657. doi: 10.1038/s41598-025-98111-y
[26]	XU D Q, WU Y Q. Improved YOLO-V3 with DenseNet for multi-scale remote sensing target detection[J]. Sensors, 2020, 20(15): 4276. doi: 10.3390/s20154276
[27]	李宗霖, 王广祥, 张立亚, 等. 基于改进YOLOv8n的煤矿带式输送异物检测研究[J]. 矿业安全与环保, 2024, 51(4): 41-48. doi: 10.19835/j.issn.1008-4495.20240604 LI Z L, WANG G X, ZHANG L Y, et al. Research on the detection of foreign objects in coal mine belt conveying based on improved YOLOv8n[J]. Mining Safety & Environmental Protection, 2024, 51(4): 41-48. (in Chinese with English abstract) doi: 10.19835/j.issn.1008-4495.20240604
[28]	苏国韶, 黄伟杰, 李政, 等. 边坡落石运动目标检测的改进YOLO模型[J]. 应用基础与工程科学学报, 2025, 33(6): 1748-1758. doi: 10.16058/j.issn.1005-0930.2025.06.016 SU G S, HUANG W J, LI Z, et al. An improved YOLO model for moving object detection of slope rockfall[J]. Journal of Basic Science and Engineering, 2025, 33(6): 1748-1758. (in Chinese with English abstract) doi: 10.16058/j.issn.1005-0930.2025.06.016
[29]	胡昊, 史天运, 关则彬. 融合混合注意力和改进YoloX的铁路落石检测方法[J]. 电子测量技术, 2022, 45(20): 110-116. HU H, SHI T Y, GUAN Z B. A railway rockfall detection method incorporating mixed attention and improved YoloX[J]. Electronic Measurement Technology, 2022, 45(20): 110-116. (in Chinese with English abstract)
[30]	BARZ B, DENZLER J. Deep learning on small datasets without pre-training using cosine loss[C]//Anon. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV). Snowmass, CO, USA: IEEE, 2020: 1360-1369.
[31]	BRIGATO L, IOCCHI L. A close look at deep learning with small data[C]//Anon. 2020 25th International Conference on Pattern Recognition (ICPR). Milan, Italy: IEEE, 2021: 2490-2497.
[32]	PAN S J, YANG Q. A survey on transfer learning[J]. IEEE Transactions on Knowledge and Data Engineering, 2010, 22(10): 1345-1359. doi: 10.1109/TKDE.2009.191
[33]	ZHUANG F Z, QI Z Y, DUAN K Y, et al. A comprehensive survey on transfer learning[J]. Proceedings of the IEEE, 2021, 109(1): 43-76. doi: 10.1109/JPROC.2020.3004555
[34]	ZAMAN A, KAMAL K, TAHIR M. Multi-camera vehicle tracking and re-identification with transfer learning for urban surveillance[J]. IEEE Transactions on Intelligent Transportation Systems, 2023, 8(24): 8234-8246.
[35]	CHEN C H, ZHANG S F, WU H T, et al. Enhancing cross-domain vehicle detection with transfer learning and source-similar sample integration[J]. Journal of Internet Technology, 2025, 26(7): 949-956. doi: 10.70003/160792642025122607010
[36]	SU G S, HUANG W J, LI C J, et al. Rockfall-YOLO deep learning model for rockfall motion object detection[J]. Journal of Computing in Civil Engineering, 2025, 39(6): 04025104. doi: 10.1061/JCCEE5.CPENG-6268
[37]	蔡鑫地, 吴四九, 何兵. 基于改进YOLOv8的道路病害检测算法[J/OL]. 公路工程: 1-14[2026-08-24]. https://link.cnki.net/urlid/43.1481.U.20260212.1849.002. CAI X D, WU S J, HE B. Road disease detection algorithm based on improved YOLOv8[J/OL]. Highway Engineering: 1-14[2026-08-24]. https://link.cnki.net/urlid/43.1481.U.20260212.1849.002. (in Chinese with English abstract)
[38]	陈鹏宇, 王烈, 梁钰墁, 等. AFL-YOLO: 基于YOLOv8改进的小目标检测算法[J]. 电讯技术, 2026, 66(2): 229-238. doi: 10.20079/j.issn.1001-893x.241024002 CHEN P Y, WANG L, LIANG Y M, et al. AFL-YOLO: An improved small object detection algorithm based on YOLOv8[J]. Telecommunication Engineering, 2026, 66(2): 229-238. (in Chinese with English abstract) doi: 10.20079/j.issn.1001-893x.241024002
[39]	刘丽丽, 王智文, 王亮, 等. 基于BiFPN和注意力机制改进YOLOv5s的车辆行人检测[J]. 现代电子技术, 2025, 48(3): 174-180. doi: 10.16652/j.issn.1004-373x.2025.03.028 LIU L L, WANG Z W, WANG L, et al. Improved YOLOv5s vehicle and pedestrian detection based on BiFPN and attention mechanism[J]. Modern Electronics Technique, 2025, 48(3): 174-180. (in Chinese with English abstract) doi: 10.16652/j.issn.1004-373x.2025.03.028
[40]	朱四超, 潘世强. 基于改进YOLO模型的无人机航拍影像滑坡目标检测应用研究[J]. 公路工程, 2025, 50(3): 213-222. doi: 10.19782/j.cnki.1674-0610.2025.03.024 ZHU S C, PAN S Q. Research on the application of landslide target detection in UAV aerial images based on improved YOLO model[J]. Highway Engineering, 2025, 50(3): 213-222. (in Chinese with English abstract) doi: 10.19782/j.cnki.1674-0610.2025.03.024