基于注意力变形和动态查询机制的交通小目标检测

Traffic small object detection based on attention deformation and dynamic query mechanism

摘要:

深度学习推动了交通目标检测发展,但复杂交通场景下密集遮挡环境中的小目标检测精度仍不足。针对上述问题提出一种注意力变形和动态查询机制的交通小目标检测算法CDAQ-DDETR,在Deformable DETR的基础上,通过引入CBAM注意力双塔机制和DCNv2可变形卷积重构原始残差网络,增强算法对密集区域交通小目标的语义获取能力；借助AFN网络思想添加低层特征,同时构建注意力感知融合金字塔模块,提高算法对多尺度中小交通目标的检测效果；依靠在原解码器前向集成动态查询机制模块结合输入图像匹配目标特性,以构建最佳查询向量提升算法对多样化背景干扰的适应泛化能力。在VisDrone2019数据集上进行实验,结果表明:CDAQ-DDETR算法在平均精确率（mAP@0.5:0.95）上已达到37.9%,在平均召回率（mAR@0.5:0.95）上已达到57.4%,相比现阶段主流SOTA算法在检测精度上提升5.5%,召回率提升8.0%,尤其针对于小目标检测精度提升6.9%,召回率提升了10.0%,同时利用可视化实验分析其更加适用于密集场景下交通小目标检测的实际应用。

Abstract:

While deep learning has advanced traffic object detection, accurately detecting small objects in complex traffic scenes with dense occlusion remains challenging. To address these issues, this paper proposes a novel small traffic object detection algorithm, CDAQ-DDETR, which incorporates an attention-based deformation and dynamic querying mechanism. Building upon Deformable DETR, the algorithm introduces the CBAM attention-based dual-tower mechanism and DCNv2 Deformable convolutions to reconstruct the original residual network, thereby enhancing the semantic acquisition capabilities for small traffic objects in dense areas. By leveraging the AFN network concept to add lower-level features and constructing an attention-aware fusion pyramid module, the algorithm improves detection performance for multi-scale small and medium traffic objects. Additionally, by integrating a dynamic query mechanism module before the original decoder, combined with matching input image characteristics, it constructs optimal query vectors, enhancing the algorithm′s adaptability and generalization ability against diverse background interferences. Experiments conducted on the VisDrone2019 dataset show that the CDAQ-DDETR algorithm has achieved a mean Average Precision (mAP@0.5:0.95) of 37.9% and a mean Average Recall (mAR@0.5:0.95) of 57.4%. Compared to the current state-of-the-art (SOTA) algorithms, there is an improvement of 5.5% in detection precision and 8.0% in recall rate, particularly, an increase of 6.9% in precision and 10.0% in recall rate for detecting small objects. Visualization experiments further demonstrate its practical applicability and superior performance in detecting small traffic objects in dense scenes.