改进AdaMixer的拥挤行人检测模型
Improved AdaMixer based crowded pedestrian detection model
林宁 1左悦 2冯兴华3
作者信息
- 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
- 折叠
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.
Abstract
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 vali-dated 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Key words
crowded pedestrian detection/DCFFN/HIPA/MHSA/multi-scale features/adaptive fusion/city persons引用本文复制引用
基金项目
广西自然科学基金(2023GXNSFAA026333)
广西教育厅中青年基金(2021KY1805)
广西教育厅中青年基金(2019KY0949)
出版年
2024
NETL
NSTL
维普
万方数据