改进YOLOv4网络的煤矿井下行人检测算法

Coal mine pedestrian detection algorithm based on improved YOLOv4 network

为满足YOLOv4网络检测井下行人的实时性要求,改进YOLOv4网络结构,将原主干特征提取网络CSPDarknet-53替换为GhostNet网络,利用ASPP结构对主干网络提取的图像特征进行增强,通过深度可分离卷积进一步减小网络参数。结果表明,改进后的网络G-A-YOLOv4能够在准确率与原网络相近的前提下,网络模型减小84 MB,检测速度提高7 fps,能够满足实时性的要求。

This paper intends to realize the real-time requirements for detecting underground pedestrians with YOLOv4 network by improving original network structure.The study involves replacing the network CSPDarknet-53 extracted from the original backbone feature with GhostNet network;enchancing the image features extracted from the backbone network with ASPP structure;and further reducing the network parameters by depth separable convolution.The results show that the improved network G-A-YOLOv4 can reduce the network model by 84 MB and increase the detection speed by 7 fps,which can meet the real-time requirements on the premise that the accuracy is similar to the original network.