摘要
针对自动驾驶汽车由于内存和计算资源对目标检测算法的限制,导致卷积神经网络在嵌入式设备上部署困难、待检测目标发生重叠或遮挡时识别困难,甚至无法识别,提出轻量级特征重用融合特征聚焦扩散金字塔的快速识别目标检测算法(FRFDFN)。首先,该算法使用特征映射冗余思想生成冗余特征图;然后,引入特征聚焦扩散金字塔网络结构,通过定制的特征聚焦与扩散机制,结合多层特征图连接,使上下文信息扩散到各检测尺度;最后,提出遮挡感知注意力检测器,并引入闭塞感知斥力损失。试验结果表明,提出的FRFDFN算法在KITTI数据集上表现良好,精确率达到97.1%、mAP50达到97.3%,参数量相较于YOLOv8n降低13.33%,检测精度却提升0.72%,兼顾低参数量与高精度要求。
Aiming at the problem that self-driving cars have difficulty in deploying convolutional neural networks on embedded devices due to the limitations of memory and computational resources on target detection algorithms,and have difficulty in recognizing the target to be detected when it overlaps or occludes,and even fail to recognize it,we propose a fast recognition target detection algorithm with lightweight feature reuse fusion feature focusing diffusion pyramid(FRFDFN).First,the algorithm uses the feature mapping redundancy idea to generate redundant feature maps;then,a feature-focused diffusion pyramid network structure is introduced to diffuse the contextual information to each detection scale through a customized feature focusing and diffusion mechanism combined with multi-layer feature graph connectivity;finally,an occlusion-aware attention detector is proposed and occlusion-aware repulsive loss is introduced.The experimental results show that the proposed FRFDFN algorithm performs well on the KITTI dataset,with an accuracy of 97.1%,a mAP50 of 97.3%,and a reduction of 13.33%in the number of parameters compared to YOLOv8n,while the detection accuracy is improved by 0.72%,which takes into account the requirements of low number of parameters and high accuracy.
作者
辛东嵘
张杰浩
张阳
XIN Dongrong;ZHANG Jiehao;ZHANG Yang(School of Civil Engineering,Fujian University of Technology,Fuzhou 350118,China;School of Transportation,Fujian University of Technology,Fuzhou 350118,China)
出处
《交通科技与经济》
2025年第3期89-96,共8页
Technology & Economy in Areas of Communications
基金
福建省自然科学基金项目(2023J01946)。
关键词
交通工程
目标检测
特征重用
特征金字塔
注意力检测器
traffic engineering
target detection
feature reuse
feature pyramid
attention detector
