1 School of Mechanical and Automotive Engineering, Guangxi University of Science and Technology, Liuzhou 545006, China
2 School of Mechanical and Automotive Engineering, South China University of Technology, Guangzhou 510640, China
Show less
文章历史+
Received
Published
Revised
2025-03-13
2026-03-08
2025-11-29
Online Date
2026-07-09
PDF (25529K)
摘要
针对山区道路落石分布杂乱无序、小目标碎石难以识别等问题,提出一种基于YOLOv8(you only look once version 8)框架的山区道路落石检测模型。该模型将PPA(parallelized patch-aware attention)模块的特征融合策略及特征提取结构进行改进,完成EFF-PPA(enhance feature fusion-PPA)模块的设计,再利用其对基础模型的主卷积模块进行优化重构,提出EFP-C2f模块,使模型保留更丰富的特征信息;根据数据集特点,在颈部网络中加入大目标检测层,提升模型对大目标的检测精度;提出一种采用部分解耦结构设计的轻量级多尺度适应性LMSA(lightweight multi-scale adaptability)检测头,该检测头由使用组归一化层的EMSConv(efficient multi-scale conv)模块和分离出组归一化层的共享参数卷积组成,在进行适当轻量化的同时,提升检测精度。在自制落石数据集中,基础模型和改进模型的平均精度分别达到87.7%和89.8%,改进模型的精度更高,并保持高检测速度,帧率达到212.1 帧/秒。在实地采集的不同场景任务中,改进模型能够准确识别各类目标,表现出良好的泛化能力。目前,关于山区道路落石检测的研究较少,本文设计中基于问题导向的模型改进思路具有很好的参考价值。
Abstract
To address the issues of disorderly distribution of rockfalls and difficult identification of small fragmented rocks on mountain roads, a rockfall detection model for mountain roads based on the YOLOv8 (you only look once version 8) framework was proposed. The feature fusion strategy and feature extraction structure of the PPA (parallelized patch-aware attention) module were improved to design the EFF-PPA (enhance feature fusion-PPA) module. This module was used to optimize and reconstruct the main convolution module of the base model, resulting in the EFP-C2f module, which retains richer feature information. Based on the characteristics of the dataset, a large-target detection layer was added to the neck network to improve the detection accuracy for large targets. A LMSA(lightweight multi-scale adaptability) detection head with a partially decoupled structure was proposed. This head consists of an EMSConv (efficient multi-scale conv) module using group normalization and a shared-parameter convolution separated from the group normalization layer. The detection accuracy was improved while appropriate lightweighting was performed. In the self-constructed rockfall dataset, the average precisions of the baseline model and the improved model are achieved at 87.7% and 89.8%, respectively. Higher accuracy is attained by the improved model while maintaining a high detection speed, with a frame rate reaching 212.1 frames per second. In different scene tasks collected in the field, various types of targets can be accurately recognized by the improved model, showing good generalization ability. Currently, limited research is conducted on rockfall detection on mountain roads, and the problem-oriented model improvement approach adopted in this design is considered to provide valuable reference.
ZHOU Hao-wei, GE Dong-yuan, YAO Xi-fan.
Multi-scale Rockfall Detection on Mountain Roads Based on Improved Parallelized Patch-aware Mechanism[J].
Science Technology and Engineering, 2026, 26(7): 2977-2989 DOI:10.12404/j.issn.1671-1815.2501714
近年来,国内外对于山区道路落石检测的研究仍处于起步阶段,相关研究很少。随着深度学习的快速发展,众多研究者利用机器视觉进行目标检测。王俊岭等[1]在YOLOv5(you only look once version 5)主干网络将浅层特征图融合到双向特征金字塔中,并增加小目标检测层,经过实验分析提升了对小目标排水管道缺陷的检测精度。Zhan等[2]基于YOLOv5提出一种边缘引导结构,将低级特征的边缘信息与深层特征融合,解决红外船舶目标小、背景复杂导致的难以识别的问题。牟家宇等[3]使用条件参数化卷积对YOLOv7-tiny的主干网络进行改进,提升对交通标志的特征提取能力。Sun等[4]基于transformer对YOLOv5的检测头进行改进,并额外增加两个更高分辨率的检测头,来提高小目标交通标志的检测精度。袁杰等[5]通过引入高效通道注意力机制提出一种苹果叶片病害检测方法,改善了YOLOv7对于不同尺度病害的特征提取能力。Chen等[6]通过引入具有可变形卷积的检测头来改进YOLOv7模型,提升模型的空间感知能力,解决物体遮挡影响检测精度的问题。Huang等[7]将通道注意力机制嵌入到YOLOv5主干网络的浅卷积中,来减少候选框中冗余信息的干扰,增强了对热轧钢带表面小缺陷的识别能力。曾勇杰等[8]基于YOLOv8引入高效多尺度卷积,通过分组进行多尺度特征提取,降低特征图冗余信息的干扰,提升检测精度。Ju等[9]针对钢板坯检测任务,通过使用分步位移卷积对YOLOv8主卷积模块进行重构,在保持精度的同时提高主干网络特征提取的效率。蒋畅江等[10]基于YOLO框架,使用尺度注意力设计动态检测头,提升模型的多尺度特征提取能力,使模型更好地适应多尺度目标的变化。因此使用深度学习网络模型结合其他改进方式来解决道路落石检测问题是一种有效的技术途径。
由于落石呈现出密集、细碎、繁杂的特性,使用传统标注方式进行标注时,任务量繁重且难以将落石全部标出,这样会导致训练效果不佳。为解决此问题,对落石的特点进行分析,按区域分类标注,将比较大或在成片、成堆石头之外的石头归为stone类别;把石头零散分布,铺成一片的石头归为a piece of stone类别;把成堆叠状态的石头归为a pile of stone类别;该标注方法能够对落石进行清晰且全面的分类标注,这样不但有助于模型的训练,并且大大削减了标注的时间。
在衡量模型改进效果时,采用多个指标进行评估。用平均精度(mean average precision,mAP)来评估模型的整体精度质量,用每秒浮点运算次数(floating point operations per second,FLOP)来评估模型的网络复杂度,以每秒帧率(frames per second,FPS)评估模型的检测速度,以参数量来评估所占内存大小,其中mAP@0.5计算表达为
WangJunling, WangChenchen, XiongYuhua. Defect detection method and effect analysis of drainage pipeline based on YOLOv5l Im[J]. Science Technology and Engineering, 2024, 24(18): 7833-7842.
[3]
ZhanW, ZhangC, GuoJ, et al. EGISD-YOLO: edge guidance network for infrared ship target detection[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2024, 17(1): 10097-10107.
MouJiayu, NanXinyuan. Real time detection algorithm for small targets in traffic signs based on YOLOv7 tiny improvement[J]. Science Technology and Engineering, 2024, 24(30): 13072-13079.
[6]
SunX, LiuK, ChenL, et al. LLTH-YOLOv5: a real-time traffic sign detection algorithm for low-light scenes[J]. Automotive Innovation, 2024, 7(1): 121-137.
YuanJie, XieLinwei, GuoXu, et al. Apple leaf disease detection method based on improved YOLOv7[J]. Journal of Agricultural Machinery, 2024, 55(11): 68-74.
[9]
ChenJ, WenR, MaL. Small object detection model for UAV aerial image based on YOLOv7[J]. Signal, Image and Video Processing, 2024, 18(3): 2695-2707.
[10]
HuangX, ZhuJ, HuoY. SSA-YOLO: an improved YOLO for hot-rolled strip steel surface defect detection[J]. IEEE Transaction on Instrumentation and Measurement, 2024, 73: 5040017.
ZengYongjie, FanBishuang, YangYawen, et al. Application of improved YOLOv8 algorithm in defect detection of wind turbine blades[J]. Journal of Electronic Measurement and Instrumentation, 2024, 38(8): 26-35.
[13]
JuH, FangY, YangH, et al. Improved lightweight YOLOv8 with DSConv and reparameterization for continuous casting slab detection on embedded device[J]. IEEE Transaction on Instrumentation and Measurement, 2024, 74: 5003712.
FengC, ZhongY, GaoY, et al. Tood: task-aligned one-stage object detection[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE, 2021: 3490-3499.
[17]
LinT, GoyalP, GirshickR, et al. Focal loss for dense object detection[J]. IEEE Transaction on Pattern Analysis and Machine Intelligence, 2017, 42(2): 2999-3007.
[18]
RedmonJ, DivvalaS, GirshickR, 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: IEEE, 2016: 779-788.
[19]
RenS, HeK, GirshickR, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transaction on Pattern Analysis and Machine Intelligence, 2016, 39(6): 1137-1149.
[20]
CaiZ, VasconcelosN. Cascade R-CNN: high quality object detection and instance segmentation[J]. IEEE Transaction on Pattern Analysis and Machine Intelligence, 2019, 43(5): 1483-1498.
[21]
XuS, ZhengS, XuW, et al. HCF-Net: hierarchical context fusion network for infrared small object detection[C]// International Conference on Multimedia and Expo. New York: IEEE, 2024: 1-6.
[22]
LüC, ZhangW, HuangH, et al. RTMDET: an empirical study of designing real-time object detectors[EB/OL]. (2022-12-16) [2024-02-10]. https://arxiv.org/abs/2212.07784.
[23]
WuY, HeK. Group normalization[C]// Proceedings of the European Conference on Computer Vision. Basel: Springer Nature, 2018: 3-19.
[24]
TianZ, ShenC, ChenH, et al. FCOS: a simple and strong anchor-free object detector[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 44: 1922-1933.
[25]
SelvarajuR R, CogswellM, DasA, et al. Grad-CAM: visual explanations from deep networks via gradient-based localization[C]// Proceedings of the IEEE International Conference on Computer Vision. New York: IEEE, 2017: 618-626.