基于道路分割的道路天气识别方法研究

Study on Method for Road Weather Recognition Based on Road Segmentation

为了实现道路场景下天气图像的准确识别,提出了一种基于道路分割的道路天气识别方法,通过设计一种结合语义分割模型提取道路区域特征的方法,构建了一种结合道路天气图像全局特征及道路特征的道路分割融合网络(RSFN)。首先,通过道路分割网络对原始图像进行预处理,获取二值化图像,并利用卷积特征掩码(CFM)获取了道路区域信息。随后,构建了由全局网络分支和道路网络分支构成的卷积神经网络,分别用于提取整体图像区域特征和专注提取道路天气特征。针对提取到的不规则道路特征,使用CFM将图像全局特征和道路局部特征进行融合。最后,通过全连接层进行了重点道路区域的天气识别,并兼顾整体的天气识别,实现了阴天、晴天、雾天、雨天和雪天5种天气的识别。通过收集多个城市地区不同路段、不同天气下的高速公路真实监控视频,构建了道路多分类天气数据集(RMWD),并与不同网络模型进行了测试结果对比。结果表明:在参数和运算量都有所降低的情况下,RSFN天气识别算法的准确率和召回率为85.40%和80.30%,分别提高了至少3.97%和3.86%;基于道路分割的道路天气识别方法将网络模型提取特征的重点区域放在了道路中,RSFN算法实现了道路天气特征的有效提取,能够有效应用于道路场景下实时准确的天气识别。

In order to realize the accurate recognition of weather images in road scenes,a road weather recognition method based on road segmentation is proposed,and a sort of road segmentation fusion network(RSFN),with the overall road weather image features and road features,is established by designing a method for extracting road area characteristics combining with the semantic segmentation model.First,the original images are preprocessed through the road segmentation network to obtain the binary images,and the road area information is obtained by using the convolutional feature mask(CFM).Subsequently,a convolutional neural network,which is composed of overall network branches and road network branches,is established and used to extract the overall image area features and focus on extracting the road weather features respectively.In view of the extracted irregular road characteristics,the overall image features and the road local features are fused with CFM.Finally,the weather recognition of key road areas is carried out through a fully connected layer,and the recognition of 5 types of weather(cloudy,sunny,foggy,rainy,and snowy)is realized considering the overall weather recognition.A road multi-class weather dataset(RMWD)is established by collecting real surveillance videos of expressway in multiple urban areas with different road sections and weather conditions,and compared with different network models on testing results.The result shows that(1)under the situation of decrease in parameters and computations,the RSFN weather recognition algorithm has accuracy rate and recall rate of 85.40%and 80.30%that improved by at least 3.97%and 3.86%respectively;(2)the key areas of extracting features from network models are placed in roads by using the road weather recognition method based on road segmentation,and the effective extraction of road weather features is realized by using RSFN algorithm,which can be applied to real-time and accurate weather recognition in road scenes effectively.