摘要
针对复杂交通场景下车辆检测存在的特征信息提取精度低、检测率低及误检漏检率高等问题,提出了一种基于改进YOLOv11的车辆检测模型FAE_V11s。首先,引入BiFormer注意力机制,通过动态稀疏注意力权重分配,增强模型对复杂场景中车辆特征的聚焦能力。其次,引入MobileNetV3网络结构替换原骨干网络,对整体网络进行轻量化改进,有效降低模型的参数量和计算量。再者,设计一种新的改进特征金字塔结构,在P2层增加256×256尺寸的特征图输出,提升P2层特征图的语义质量与细节保留能力,实现更高效、精准的目标检测。最后,设计FPv2_IoU损失函数,解决因惩罚因子而引发的锚框膨胀问题及训练样本分布不均衡的问题,进而有效加速模型收敛。在UA-DETRAC数据集和自建数据集的实验中,与YOLOv11n模型相比,FAE_V11s参数量减少了约3.64%,mAP@0.5和mAP@0.50∶0.95分别提高了13.29和11.13个百分点,整体表现出良好的检测性能。
Aiming at the problems of low feature information extraction accuracy,low detection rate and high false detection and leakage rate in complex traffic scenarios,an improved vehicle detection model FAE_V11s based on YOLOv11 is proposed.Firstly,the BiFormer attention mechanism is added to enhance the target detection capability,optimize the computational efficiency,and improve the model’s generalization ability.Secondly,the MobileNetV3 net-work structure is introduced to replace the original backbone network,making the overall network more lightweight and effectively reducing the model’s parameter quantity and computational load.Furthermore,a new improved feature pyr-amid structure is proposed,adding a 256×256-sized feature map output at the P2 layer to enhance the semantic quali-ty and detail retention ability of the P2 layer feature map,achieving more efficient and accurate target detection.Final-ly,the FPv2_IoU loss function is proposed to solve the anchor box expansion problem caused by the penalty factor and the imbalance of training sample distribution,thereby effectively accelerating the model convergence.In experiments on the UA-DETRAC dataset and the self-built dataset,compared with the YOLOv11n model,FAE_V11s reduces the parameter quantity by approximately 3.64%,and the mAP@0.5 and mAP@0.50∶0.95 are increased by 13.29 and 11.13 percentage points respectively,demonstrating excellent detection performance overall.
作者
宋晓茹
王嘉乐
刘通
连扬志
SONG Xiaoru;WANG Jiale;LIU Tong;LIAN Yangzhi(School of Electronic Information Engineering,Xi’an University of Technology,Xi’an 710021,China;Science and Technology Laboratory on Electro-mechanical Dynamic Control,Xi’an 710065,China)
出处
《激光杂志》
北大核心
2025年第8期65-73,共9页
Laser Journal
基金
国防科技重点实验室稳定支持课题(No.6142601012306)
陕西省重点研发计划项目(No.2021GY287)。
