基于改进MobileNet-YOLOv3级联模型的内涝及受灾个体监测研究

Monitoring of waterlogging and affected individuals based on improved MobileNet-YOLOv3 cascade model

为解决现有城市内涝监测技术机动性低、设备自身缺陷、光照不均以及其他噪声因素引起的识别精度低的问题,建立改进的MobileNet-YOLOv3级联模型。首先,利用圆形LBP算法改进MobileNet模型实现监控视频内涝自动识别;改进的MobileNet模型引进圆形LBP算子提取特征,将MobiletNet卷积提取的特征与圆形LBP算子提取的特征融合,融合后的特征经过MobileNet主体网络学习训练得到最终的特征后采用全连接层分类输出监测结果;其次,通过改进YOLOv3算法实现对内涝点车辆、行人等受灾个体视频的智能识别,改进的YOLOv3算法将模型输入改为MobileNet模型提取出的特征,并且引入CIOU损失函数;最后,将两改进算法级联实现完整的内涝及受灾个体监测功能。研究结果表明:改进模型整体识别准确率达到90%以上,实现对多个特征的融合应用,使各特征之间联系更加密切,能较为准确地对城市道路或隧道等场景进行积水监测及受灾个体识别。研究结果可为内涝灾害预防提供技术支持。

In order to solve the problem of low recognition accuracy caused by low mobility,equipment defects,uneven illumination and other noise factors of existing urban waterlogging monitoring technologies,an improved MobileNet-YOLOv3 cascade model was established.Firstly,the circular LBP algorithm was used to improve the MobileNet model to realize the automatic identification of waterlogging in surveillance video.In the improved MobileNet model,the circular LBP operators were introduced to extract features,and the features extracted by MobletNet convolution and circular LBP operators were fused,and together through the learning and training of MobileNet subject network,the single-layer full-connection layer classification was adopted.Secondly,the YOLOv3 algorithm was improved to realize the intelligent identification on the video of the affected individuals such as vehicles and pedestrians at the waterlogging point.The improved YOLOv3 algorithm changed the model input to the features extracted from the MobileNet model,and introduced into the CIOU loss function.Finally,the two improved algorithms were cascaded to realize the complete monitoring function of waterlogging and affected individuals.The research results show that the overall identification accuracy of the improved model is more than 90%,the integration of multiple features is realized,the relationship between various features is closer,and the water monitoring and the identification of affected individuals in urban roads,tunnels and other scenes can be more accurate.The results can provide technical support for the waterlogging disaster prevention.