基于注意力机制的煤矿井下行人检测的轻量化网络结构

Lightweight network structure of underground pedestrian detection in coal mine based on attention mechanism

针对井下监控成像模糊,行人检测准确率低及检测算法在小型设备中运行速度慢的问题,将ECA注意力机制引入到YOLOv4-tiny网络的CSPdarknet53-tiny尾部,通过改进轻量型网络YOLOv4-tiny,以及不降维的局部跨信道交互策略和自适应选择一维卷积核大小的方法,提高网络的检测性能。结果表明,改进YOLOv4-tiny网络训练速度快于YOLOv3网络,准确率比YOLOv4-tiny和SSD300网络分别提高了2.69%和4.25%。

This paper proposes a novel underground pedestrian detection method designed to address the problems such as fuzzy imaging in underground monitoring,lower accuracy of underground pedestrian detection,and slower operation affecting underground pedestrian detection algorithm in small equipment.The method is realized by improving the YOLOv4-tiny network,a lightweight network of YOLOv4,introducing ECA attention mechanism into the tail of CSPdarknet53-tiny in YOLOv4-tiny network,and improving the performance of the network using the local cross-channel interaction strategy without dimension reduction and the method of adaptively selecting the size of one-dimensional convolution kernel.The results show that the improved YOLOv4-tiny network has a faster training speed than YOLOv3 network,enabling the increased accuracy rate by 2.69%and 4.25%respectively compared with YOLOv4-tiny network and SSD300 network.