Anomaly Detection In Urban Expressway Using Semantic Segmentation

Scene parsing is a hot research area in the field of computer vision.This task hopes that the computer has a global understanding of the entire picture rather than a partial one,which increases the need of computers for understanding the details of the image.Semantic segmentation is a key technology for scene parsing.Semantic segmentation is to determine the class of each pixel in the image,divide adjacent pixels with the same class into blocks,and finally divide the picture into multiple parts.For road scenes,the application of semantic segmentation has many difficulties.On the one hand,semantic segmentation itself will cause problems such as discontinuity in the segmentation results due to the existence of the upsampling network.On the other hand,for practical reasons,road pictures often have a problem with a large depth of field,which will cause objects misclassification of the same category in the near and distant scenes.In response to the above problems,we propose our model to obtain better segmentation results.The main contributions of our work are as follows:First,we propose a feature pyramid module based on edge detection.Edge detection can improve the sensitivity of the entire network to edge information and improve the extraction ability of the segmentation model for semantic segmentation.And the feature pyramid can make the model more effective in the recognition of small objects.During training,the module directly uses edge detection operators to calculate the convolution feature map in the neural network.Then it merges the obtained edge detection features of different scales through the feature pyramid,which will allow the model to get more Features that are conducive to smooth edges.Second,this paper establishes a complete semantic segmentation model and designs the corresponding loss function for the model.This paper divides the loss function into two parts,the first part is the loss function of the semantic segmentation model,and the second part is the loss function of the edge feature pyramid.The encoder and decoder in the model are respectively connected to an edge feature pyramid.The output feature representation by the two should have similarity due to the symmetry structure of the model.Because of the similarity.the upsampling feature representation should be as close as possible to the downsampling feature expression.The similarity of the two is compared by MSE loss function,and the difference is put into the total loss function as the loss function to supervise the model training.Third,a variety of anomaly road traffic and corresponding detection standards are defined from the perspective of semantic segmentation for the actual scene of Jingtong Expressway.And we qualitatively evaluated the detection performance of semantic segmentation on the traffic situation of Jingtong Expressway from the perspective of visualization.In addition,the public dataset is used to quantitatively evaluate the effect of road condition anomaly detection by evaluating the segmentation accuracy of semantic segmentation.The experiment on Cityscapes datasets demonstrates that the accuracy is improved by 9.1%over the SegNet.