基于多层特征融合的多光谱行人检测方法

Multi-spectral pedestrian detection based on multi-layer fusion features

针对行人检测模型在全天候场景下特征提取能力差、检测准确率低的问题,提出一种基于多层特征融合的多光谱行人检测方法。设计一种基于通道注意力机制的可见光与红外特征的融合方式,提升模型的特征融合效率;采用多层融合特征设计特征金字塔网络,提升模型的特征提取能力;引入自适应特征融合机制对检测层进行特征图尺度调整,降低尺度冲突对模型性能的影响。在KAIST数据集上进行实验,其结果表明,模型的检测性能有一定提升。

For the problems that the pedestrian detection model has poor feature extraction ability and the low detection accuracy in all-reasonable scenes,a multi-spectral pedestrian detection method based on multi-layer fusion features was proposed.A fusion method of visible light features and infrared features based on the channel attention mechanism was designed to improve the fusion efficiency of feature maps.The multi-layers fusion features were used to design the feature pyramid network to improve the feature extraction ability of the model.An adaptively spatial feature fusion mechanism was used to adjust the feature map’s scale of the detection layers,so the impact of scale conflict on the accuracy of the model detection was reduced.It is verified on the KAIST dataset,the experimental results show that the performance of this method is better than that of those existing multi-spectral pedestrian detection methods.