Research On Road Anomaly Detection Based On Semantic Segmentation

Semantic segmentation is a classic topic in computer vision.The task is to assign each pixel of the image a pre-defined class label.Semantic segmentation has wideranging applications,such as scene parsing,autonomous driving and medical segmentation,to name a few.In the field of transportation,semantic segmentation is meaningful to abnormal event recognition based on surveillance video with image processing.In this paper,we analyze multiple mechanisms about advanced nervous networks and summarize research results of semantic segmentation.Lots of Convolution neural networks based on Fully Convolution Network have achieved impressive results on semantic segmentation task.There are still many novel structures and robust components which are devoted to explore potentials of features for more excellent performance and faster inference speed.1.Pyramid Pooling Module with SE1 Cblock and D2 SUpsample Network based on convolution neural network is designed,and it’s starts from Pyramid Scene Parsing Network.Two powerful modules focusing on attention and upsampling mechanisms respectively are embedded in model to improve the accuracy of the deep convolution neural network in semantic segmentation task.2.Spatial and Parallel Context Combined Network focuses on faster inference speed without sacrificing accuracy.By using well-designed structures such as parallel branch based on Res Net-18 to extract semantic features and light-weighted branch to extract spatial features and feature fusion module belonging to BiSeNet,the real-time semantic segmentation task can be accurately and efficiently completed.3.Abnormal event recognition task under traffic scene is solved by optimized SPCCNet and image processing technology.