摘要
由于道路交通场景存在许多遮挡和小目标物体,很容易出现误检测和漏检,因此提出一种基于RT-DETR的改进目标检测模型来提升检测性能。在特征提取网络方面,采用经过Shuffle Attention(SA)注意力机制增强的ResNet-18,用来加强网络特征提取能力;同时引入Cascaded Group Attention(CGA)机制替换原模型尺度内特征交互(AIFI)模块中的多头自注意力机制(MHSA),成功减少计算冗余,提升了模型性能。最后,构建专门针对道路交通场景的数据集进行实验。模型在RTX4070ti GPU平台上进行了性能验证。性能评估表明,改进后的模型在平均精度(mAP)上达到72.9%,较原RT-DETR模型提升2.1%。此外,在每秒帧数(FPS)方面,改进模型同样表现出色、达到132.1,优于RT-DETR 9帧和YOLOv8m 23帧。综合实验结果显示,本研究提出的改进模型不仅保持了高检测精度,还成功地加速了模型计算。这些改进对于实时且精确处理道路交通场景的目标检测具有重要的实用价值。
Due to the prevalence of occlusions and small targets in road traffic scenarios,there is a high propensity for false detections and omissions.Consequently,the paper proposes an enhanced object detection model based on RT-DETR.In the realm of feature extraction,the proposed model employs a ResNet-18 framework augmented with a Shuffle Attention(SA)mechanism,bolstering its feature extraction capacity.Additionally,the research integrates a Cascaded Group Attention(CGA)mechanism,substituting the Multi-Head Self-Attention(MHSA)within the original model′s Attention Intra-Feature Interactions(AIFI)module,thereby significantly reducing computational redundancy and enhancing model performance.Experiments are conducted using a dataset specifically designed for road traffic scenes.The model′s performance is validated on an RTX4070ti GPU platform.Performance evaluation reveals that the improved model achieves a mean Average Precision(mAP)of 72.9%,marking a 2.1%increase compared to the original RT-DETR model.Moreover,in terms of Frames Per Second(FPS),the enhanced model reaches 132.1 FPS,surpassing RT-DETR by 9 frames and YOLOv8m by 23 frames.The comprehensive experimental results demonstrate that the proposed improved model not only maintains high detection accuracy,but also significantly accelerates model computation.These advancements hold significant practical value for real-time and precise object detection in road traffic scenarios,leveraging the capabilities of deep learning in object detection.
作者
唐坤俊
宁媛
刘聂天和
TANG Kunjun;NING Yuan;LIUNIE Tianhe(School of Electrical Engineering,Guizhou University,Guiyang 550025,China;Guiyang Huaxi Power Supply Bureau of Guizhou Power Grid Co.,Ltd.,Guiyang 550025,China)
出处
《智能计算机与应用》
2025年第9期82-89,共8页
Intelligent Computer and Applications
基金
贵州省科技计划基金(黔科合ZK2022135)。
