结合通道注意力的无人机遥感影像行人检测方法

UAV Remote Sensing Image Pedestrian Detection Method Combined with Channel Attention

在无人机遥感影像行人检测任务中,针对无人机影像背景复杂、行人目标尺度多样化且样本个数稀少等问题,提出一种结合通道注意力的单阶段卷积网络模型。模型以并联卷积核构成特征提取层,同时引入通道注意力机制来强化模型对行人特征的学习能力;在特征强化网络中进一步通过融合跨尺度特征,构建了4个输出层的特征强化网络,最后使用增广后的训练集完成模型的训练。实验结果表明,本文所提出的行人检测模型能够对曝光、阴影等复杂效果下以及不同角度、不同距离下的行人目标实施稳定检出,在测试环境下的检测速度也达到实时检测水平。

In the task of pedestrian detection in UAV remote sensing images,a single-stage convolutional network model combined with channel attention is proposed to solve the problems of complex UAV image background,diverse pedestrian target scales and few samples.The model uses parallel convolution kernels to form a feature extraction layer,and at the same time introduces a channel at-tention mechanism to enhance the model's ability to learn pedestrian features;in the feature enhancement network,a feature enhance-ment network of four output layers is further constructed by fusing cross-scale features.Finally,use the augmented training set to complete the training of the model.The experimental results show that the pedestrian detection model proposed in this paper can stably detect pedestrian targets under complex effects such as exposure and shadows,as well as at different angles and distances,and the de-tection speed in the test environment also reaches the level of real-time detection.