改进AdaMixer的拥挤行人检测模型  

Improved AdaMixer based crowded pedestrian detection model

作　　者：林宁 左悦 冯兴华 LIN Ning;ZUO Yue;FENG Xing-hua(School of Information Engineering,Nanning University,Nanning 530200,China;School of Civil and Architectural Engineering,Nanning University,Nanning 530200,China;School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China)

机构地区：[1]南宁学院信息工程学院,广西南宁530200 [2]南宁学院土木与建筑工程学院,广西南宁530200 [3]西南科技大学信息工程学院,四川绵阳621010

出　　处：《计算机工程与设计》2024年第5期1525-1532,共8页Computer Engineering and Design

基　　金：广西自然科学基金项目(2023GXNSFAA026333);广西教育厅中青年基金项目(2021KY1805、2019KY0949)。

摘　　要：为应对拥挤场景中行人检测的挑战,提出一种改进型AdaMixer的行人检测模型。提出一种头交互位置感知多头自注意力机制(HIPA-MHSA),以加强对相互遮挡目标的辨别能力。采用深度卷积前馈网络(DCFFN),进一步提升模型的特征提取效能。通过在公开数据集上进行验证,表明了所提出模型的有效性。填补了现有查询型目标检测器在语义和空间自适应方面的不足,提高了拥挤场景下行人检测的精度,表现优于其它对比模型。To face the challenge of pedestrian detection in crowded scenes,an enhanced AdaMixer-based pedestrian detection model was proposed.The head-interaction position-aware multi-head self-attention mechanism(HIPA-MHSA)was introduced to enhance the discriminative capability for mutually occluded targets.A deep convolutional feedforward network(DCFFN)was employed to further improve the model’s feature extraction efficiency.The proposed model’s effectiveness was thoroughly validated on publicly available datasets.The shortcomings of existing query-based object detectors in semantic and spatial adaptation are addressed,with the pedestrian detection accuracy in crowded scenarios being enhanced and other comparative models being outperformed.

关 键 词：拥挤行人检测 深度卷积前馈网络 头交互位置感知 多头自注意力机制 多尺度特征 自适应融合 城市行人 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]