Semantic Segmentation Of Traffic Scenes Based On Category Grouping Under Abnormal Weather Condition

In recent years,with the rapid development of the field of computer vision,autonomous driving has become a research hotspot.Traffic scene semantic segmentation mainly processes the data of driving scenes and is an important technical support in the field of autonomous driving.However,the current traffic scene semantic segmentation is basically researched on normal condition data,and the common and more difficult abnormal weather conditions are not considered.Therefore,studying the traffic scene semantic segmentation under abnormal weather conditions is a very challenging task,which is of great significance to the field of autonomous driving.Aiming at the task of autonomous driving,this paper proposes a new idea to solve the semantic segmentation of traffic scenes under abnormal weather.The main research contents are as follows:(1)The idea of grouping and segmenting object classes according to importance is proposed in this paper,and a multi-head network based on class grouping is constructed.This network realizes the separate segmentation task of each group by setting different header decoding networks for different groups.Later,in the model training process,different weights are applied to different groups according to their importance,so as to achieve the purpose of giving priority to the segmentation results of important groups.Simultaneously,according to the characteristics of the network structure,a class relation module is proposed in this paper,and the effect is improved by encoding the correlation information between classes in advance.Finally,On the MIo U evaluation metric,compared with the corresponding normal semantic segmentation model,on the the model based on class grouping leads 12% in 100 mm intensity simulated rainy day data,5.1% in 75 m visibility simulated fog day data,about 5% on average in ACDC data set and 2.63% in abnormal data partitioned by bdd100 k data set.(2)Further improvements are proposed for the method based on class grouping.First,an improved FW loss function based on Focal Loss is proposed in this paper to achieve the effect of applying different weights to different groups.After combining the class grouping model with the loss function,it shows a certain improvement effect on 4 data sets.Secondly,the method based on class grouping and the method of abnormal data preprocessing are compared experimentally in this paper,and the characteristics of the two methods are obtained.And a combined processing method with the characteristics of these two methods is obtained by combinaing them.Finally,by combining the effective rain and fog preprocessing algorithm,the class grouping model obtains better results in each group segmentation’s effect and overall segmentation’s effect compared with the normal segmentation model.