基于RSG-YOLOv10n道路场景抗干扰与小目标检测方法
孔飞一，付振山*，王玉刚，付聪，戴先鑫，马栋
山东交通学院船舶与港口工程学院，山东 威海 264209
摘要：针对自动驾驶道路场景目标检测存在背景干扰、远距离小目标、特征缺失等问题，以YOLOv10n(you only look once version 10n)模型为基准，引入感受视野卷积（receptive field attention convolution, RFAConv）模块替换Conv模块，提取多尺度感受野空间特征，结合注意力机制动态分配权重，增强模型复杂图像处理能力；引入小目标增强金字塔（small object enhance pyramid, SOEP）模块，采用改进的CSP-OmniKernel(cross stage partial-omniKernel)模块进行特征信息整合，显著提高小目标检测性能；引入全局通道空间注意力（global channel-spatial attention，GCSA）模块，通过通道注意力、通道洗牌与空间注意力的耦合机制强化特征图表达，捕捉特征图全局依赖关系，增强特征提取能力，形成RSG-YOLOv10n（RFAConv-SOEP-GCSA-YOLOv10n）复杂道路场景小目标检测模型，开展消融试验、模型性能对比试验、通用性验证试验和可视化检测效果试验。试验结果表明：引入RFAConv、SOEP、GCSA 3个模块形成的RSC-YOLOv10n模型的精确率P、召回率R、交并比阈值为50时的平均精度均值EmAP50、交并比阈值从50增至95（步长5）时的平均精度均值EmAP50-95比YOLOv10n模型分别增大6.3百分点、2.5百分点、3.5百分点和2.8百分点，检测精度明显提高；与较低参数量模型（YOLOv5n、YOLOv8n、YOLOv10n）、相近参数量模型（YOLOv3-ting、YOLOv6n、YOLOv7-ting）及较高参数量模型（RT-DETR-L）相比，RSC-YOLOv10n模型在较低参数量和浮点运算次数基础上，检测精度最高；RSC-YOLOv10n模型对由沙尘、大雪、浓雾等多种恶劣气象图片组成的通用性验证数据集进行可视化检测，其EmAP50、EmAP50-95比YOLOv10n模型分别增大0.9百分点、1.3百分点，鲁棒性和泛化能力较高；RSG-YOLOv10n模型在可视化检测效果试验中对密集遮挡、复杂光照、恶劣天气及远近目标共存等道路场景的特征提取与小目标检测能力较强。
关键词：小目标检测；背景干扰；特征缺失；RSG-YOLOv10n；注意力机制
Anti-interference and small object detection method for road scene based on RSG-YOLOv10n
KONG Feiyi, FU Zhenshan*, WANG Yugang, FU Cong, DAI Xianxin, MA Dong
Naval Architecture and Port Engineering College, Shandong Jiaotong University, Weihai 264209, China
Abstract: Addressing the issues of background interference, distant small objects, and feature loss in autonomous driving road scene object detection, an improved model based on You Only Look Once version 10n (YOLOv10n) is proposed. The receptive field attention convolution (RFAConv) module is introduced to replace the Conv module, extracting multi-scale receptive field spatial features and dynamically allocating weights through attention mechanisms to enhance the model′s complex image processing capabilities. A small object enhance pyramid (SOEP) module is incorporated, utilizing an improved cross stage partial-omniKernel (CSP-OmniKernel) module for feature information integration, significantly improving small object detection performance. The global channel-spatial attention (GCSA) module is introduced to enhance feature map representation through the coupling mechanism of channel attention, channel shuffle, and spatial attention, capturing global dependencies in feature maps and enhancing feature extraction capabilities, forming the RSG-YOLOv10n complex road scene small object detection model. Ablation experiment, model performance comparison experiment, generalization validation experiment, and visualization detection effect experiment are conducted. Experimental results show that after introducing the RFAConv, SOEP, and GCSA modules, the RSG-YOLOv10n model′s precisionP, recallR, mean average precision at 50 intersection over union thresholdEmAP50, and mean average precision averaged over intersection over union thresholds from 50 to 95 (with a step of 5)EmAP50-95is improved by 6.3 percentage points, 2.5 percentage points, 3.5 percentage points, and 2.8 percentage points respectively compared to the YOLOv10n model, with significantly enhances detection accuracy; compared with lower parameter models(YOLOv5n, YOLOv8n, YOLOv10n), similar parameter models(YOLOv3-tiny, YOLOv6n, YOLOv7-tiny), and higher parameter model(RT-DETR-L), the RSG-YOLOv10n model achieves the highest detection accuracy with lower parameter count and floating-point operations; the RSG-YOLOv10n model performs road scene object detection on a generalization validation dataset composed of various adverse weather images including sandstorms, heavy snow, and dense fog, with EmAP50 and EmAP50-95 increasing by 0.9 percentage points and 1.3 percentage points respectively compared to the YOLOv10n model, demonstrating high robustness and generalization capability; the RSG-YOLOv10n model exhibits strong feature extraction and small object detection capabilities in visualization detection experiments for road scenes with dense occlusion, complex lighting, adverse weather, and coexisting near and distant objects.
Keywords: small object detectionn; background interference; feature loss; RSG-YOLOv10n; attention mechanism
