基于L-ConvNeXt网络的变电站人员安全操作检测方法

Substation Personnel Safety Operation Detection Based on L-ConvNeXt Network

针对变电站复杂环境中作业人员操作检测网络参数量过大、作业人员部分特征不明显等问题,本文提出一种基于L-ConvNeXt网络的变电站人员检测方法。首先,网络的主干特征提取部分由轻量化ConvNeXt模块搭建而成,保证网络特征提取能力同时使主干部分保持较低的参数量;其次,选择TPH(Transformer prediction head)作为网络末端检测头,加强网络对低分辨率特征的检测;最后,引用VariFocal Loss作为目标损失函数中的分类损失和置信度损失,进一步增加网络对正样本的损失权重。在天池公共数据集上的实验结果表明:本文网络模型获得较好的检测效果,其平均检测精度达到89.6%,模型参数量为13.2×10^(6),能够有效地检测变电站人员的作业情况,满足变电站复杂场景下的检测需求。

A substation personnel detection based on the L-ConvNeXt network is proposed to solve the problems of excessive amount of network parameters and unclear characteristics of operational personnel detection in the complex environment of substations.Firstly,the backbone feature extraction network is built by the lightweight ConvNeXt to enhance the feature extraction capability of the network while keeping the backbone with a low parameter.Then,the transformer prediction head(TPH)is selected as the end detection head to enhance the detection of low-resolution features by the network.Finally,VariFocal Loss is introduced to replace Focal Loss as target loss and confidence loss to improve the loss weight of the network for positive samples.The experimental results on the public dataset of Tianchi show that the proposed network achieves a better detection results,the average detection accuracy reaches 89.6%and the number of model parameters is 13.2×10^(6),which can effectively detect the operation of substation personnel and meet the detection requirements of substation operators in complex scenarios.