基于通道注意力的轻量行人检测算法

Lightweight pedestrian detection algorithm based on channel attention

YOLOv3算法参数量和计算量较大,不适合在移动端上使用。针对这个问题,文中通过优化YOLOv3算法提出一种基于注意力机制的轻量行人检测算法。首先,采用轻量级网络优化YOLOv3模型结构,减少模型的参数量和计算量;其次,设计下采样通道注意力模块代替Darknet53中的下采样层;最后,为了进一步丰富目标特征信息,增强小尺度行人的检测能力,引入特征增强模块。在INRIA数据集上的实验结果表明,所提出方法参数量相比YOLOv3模型降低约18,模型平均准确率提高3.85%。相比其他轻量化算法,提出的算法模型复杂度更低并且检测性能更好。

The YOLOv3 algorithm is not suitable for use on mobile terminals due to its parameter quantity and calculation quantity are large.On this basis,a lightweight pedestrian detection algorithm based on attention mechanism is proposed by optimizing the YOLOv3 algorithm.The YOLOv3 model structure is optimized by means of lightweight network to decrease the parameters quantity and calculated quantity of model.A down⁃sampling channel attention module is designed to replace the down⁃sampling layer in Darknet53.The feature enhancement module is introduced to further enrich the target feature information and improve the detection ability of small⁃scale pedestrians.The experimental results on the INRIA datasets show that the proposed method can reduce parameters count by 18 in comparison with the YOLOv3 model,and increase the average accuracy by 3.85%.In comparison with other lightweight algorithms,the proposed algorithm has lower complexity and better detection performance.