交通场景目标检测算法研究与改进——基于改进YOLOv11的小目标与遮挡目标检测方法

Research and Improvement of Object Detection Algorithms in Traffic Scenarios—A Small and Occluded Object Detection Method Based on Improved YOLOv11

摘要: 针对YOLOv11在交通场景目标检测中特征提取不充分、多尺度目标适应性差、小目标漏检率高及模型计算量大等问题，提出一种融合坐标注意力机制与轻量化特征金字塔的改进YOLOv11算法。在主干网络引入轻量化坐标注意力模块，强化交通目标关键区域空间感知；重构颈部网络为改进型BiFPN结构，结合跨阶段局部连接与深度可分离卷积实现高效多尺度特征融合；优化检测头动态锚框生成策略与标签分配方式，采用动态IoU阈值及自适应正负样本选择；重构损失函数，设计自适应焦点损失与改进型EIoU损失分别解决小目标检测与边界框回归问题。在自建UrbanTraffic-2026数据集及公开数据集VisDrone、UA-DETRAC上的实验结果表明，改进算法的mAP@0.5分别提升3.2%、2.8%和4.1%，小目标检测mAP_s最高提升8.9个百分点，推理速度保持在45 FPS以上，参数量与计算量分别降低7.9%和14.6%，在保证检测精度的同时实现模型轻量化，满足智能交通系统实时检测需求。

Abstract: To address the issues of insufficient feature extraction, poor adaptability to multi-scale objects, high false negative rate for small objects, and large computational cost of YOLOv11 in traffic scene object detection, an improved YOLOv11 algorithm that integrates a coordinate attention mechanism and a lightweight feature pyramid is proposed. A lightweight coordinate attention mechanism is introduced into the backbone network to enhance the spatial position perception ability of small objects and key regions. The neck network is reconstructed with an improved lightweight BiFPN structure, combined with cross-stage partial connections and depthwise separable convolutions. This design improves the fusion effect of multi-scale fine-grained features. The dynamic anchor box generation strategy and label assignment method of the detection head are optimized, and a dynamic IoU threshold and adaptive positive and negative sample selection are adopted. The loss function is reconstructed, and an adaptive focus loss and an improved EIoU loss are designed to solve the small object detection and bounding box regression problems, respectively. Experimental results on the self-built UrbanTraffic-2026 dataset and the public datasets VisDrone and UA-DETRAC show that the improved algorithm improves the mAP@0.5 by 3.2%, 2.8%, and 4.1%, respectively, and the mAPs of small object detection by up to 8.9 percentage points. The inference speed remains above 45 FPS, and the number of parameters and computational cost are reduced by 7.9% and 14.6%, respectively. The model is lightweight while maintaining detection accuracy, meeting the real-time detection requirements of intelligent transportation systems.