改进的卷积神经网络在地铁保护区无人机安全巡查中的应用

Application of improved convolutional neural network in UAV safety patrol in subway protection area

本文基于低空无人机遥感数据,应用人工智能和机器学习技术的卷积神经网络(CNN)算法,通过相应的优化改进,自动提取地铁保护区内施工工程位置,为地铁保护区内安全风险源的识别和精准定位提供了一种新的技术方法。本文首先评估了实验数据情况,通过数据样本集增强处理,改善了小目标特征区分难度大及识别过程中空间层级化信息和小物体信息损失严重的问题。其次,应用未改进的卷积神经网络你只需看一次5版(YOLOV5)进行实验,结果表明未改进的YOLOV5卷积神经网络在本实验数据集中训练效果不佳,检测精度低,存在错检漏检现象。最后,针对实验凸显的问题,对YOLOV5卷积神经网络进行改进,引入基于网络的迁移学习、增加注意力机制模块等优化措施,提高了风险源目标的检测效率,解决了信息超载的问题。优化改进后的YOLOV5卷积神经网络平均检测精度超过92.3%,实现了地铁保护区内安全风险源的自动化精准识别、快速定位、位置解码以及信息汇总的全面巡查。

Based on the remote sensing data of low-altitude unmanned aerial vehicles(UAVs),this paper applied the convolutional neural network(CNN)algorithm combining artificial intelligence and machine learning technology to automatically extract the construction project location in the subway protection area through corresponding optimization and provided a new technical method for the identification and accurate positioning of safety risk sources in the subway protection area.Firstly,this paper evaluated the experimental data and solved the problems of distinguishing small target features and reducing serious loss of spatial hierarchical information and small object information in the identification process through data sample set enhancement processing.Secondly,the unimproved CNN,namely you only look once version(YOLOV5)was used for experiments.The results show that the training effect of YOLOV5 in the experimental dataset is not good,and the detection accuracy is low.In addition,there is a phenomenon of wrong detection and missed detection.Finally,in view of the problems highlighted in the experiment,YOLOV5 is improved,and optimization measures such as network-based transfer learning and attention mechanism module are introduced to improve the detection efficiency of risk source targets and solve the problem of information overload.The average detection accuracy of the optimized YOLOV5 reaches more than 92.3%,and the comprehensive patrol involving automatic accurate identification,rapid positioning,location decoding,and information aggregation of safety risk sources in the subway protection area is realized.