Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (9): 329-336.DOI: 10.3778/j.issn.1002-8331.2111-0264

• Engineering and Applications • Previous Articles    

3D Object Detection in Substation Scene Based on Graph Neural Network

ZHANG Ting, ZHANG Xingzhong, WANG Huimin, YANG Gang, WANG Dawei   

  1. 1.College of Software, Taiyuan University of Technology, Jinzhong, Shanxi 030600, China
    2.Electric Power Research Institute of State Grid Shanxi Electric Power Company, Jinzhong, Shanxi 030600, China
  • Online:2023-05-01 Published:2023-05-01



  1. 1.太原理工大学 软件学院,山西 晋中 030600
    2.国网山西省电力公司 电力科学研究院,山西 晋中 030600

Abstract: In the three-dimensional scene of a substation,the precise positioning and identification of inspectors and live equipment is a prerequisite for improving the level of personnel safety management and control. Aiming at the problem of inaccurate target positioning and recognition in complex scenes of substations, a method of 3D object detection in substation scene based on graph neural network is proposed. The method is designed based on the point-GNN structure. In the vertex feature extraction stage, the PCS(point-channel-sphere) attention structure is proposed to extract more abundant key point feature information. In the GNN edge feature aggregation stage, an overall pooling mechanism is adopted to take into account the maximum pooling and the mean pooling to obtain richer global features, improving the model loss function, using Focal Loss as the classification loss to make the training pay more attention to the previous scenic spots, and using DIoU Loss as the regression loss to make the regression task more efficient. Training and testing on the self-built substation scene dataset, experiments show that the mAP value of this method reaches 73.81%, which is better than the benchmark model. It can improve the detection effect of objects in the substation scene and has certain practical value for improving the level of personnel safety management and control.

Key words: graph neural network, 3D object detection, point cloud, attention, overall pooling, loss function

摘要: 在变电站三维场景中,对巡检人员和带电设备的精确定位与识别是提高人员安全管控水平的前提。针对变电站复杂场景中目标定位与识别不准的问题,提出了一种基于图神经网络的变电站场景三维目标检测方法。该方法基于point-GNN结构设计,在顶点特征提取阶段,提出PCS(point-channel-sphere)注意力结构,提取更加丰富的关键点特征信息;在GNN边缘特征聚合阶段,采用统筹性池化机制,兼顾最大池化和均值池化从而获取更丰富的全局特征;改进模型损失函数,将Focal Loss作为分类损失使训练更加关注前景点,将DIoU Loss作为回归损失使回归任务更高效。在自建的变电站场景数据集上进行训练与测试,实验表明该方法mAP值达到73.81%,优于基准模型,能够改善变电站场景中目标的检测效果,对提高人员安全管控水平具有一定的实用价值。

关键词: 图神经网络, 三维目标检测, 点云, 注意力, 统筹性池化, 损失函数