Research On Semantic Understanding Algorithms Of Traffic Scene

With the rapid development of Intelligent Transportation System(ITS),the performance of driverless vehicles is improving day by day.Scene understanding technology is one of the important technologies of driverless vehicles,and semantic segmentation is the first and key step of scene understanding technology.Influenced by complex traffic environment,the result of semantic segmentation is prone to the problems of poor processing of low illumination edges and blurred details of segmentation objects.In the phase of traffic image acquisition,it is extremely difficult to label data sets manually.To solve these problems,this paper studies the semantic segmentation algorithm based on deep convolution neural network as follows.Aiming at the problem of poor processing of low illumination edges in image semantics segmentation,this paper firstly uses deep residual network to learn more high-order semantics features in image;secondly,uses regional candidate network to accelerate the generation of candidate region blocks of target to be segmented;secondly,designs fusion algorithm to fuse candidate region blocks,and eliminates duplicate candidate regions.Then,the low illumination edge is searched in the fused target area block,and the low illumination edge feature is enhanced by using the local enhancement algorithm with less distortion cost.In order to solve the problem of blurred details in image segmentation,a semantic segmentation algorithm based on full convolution network is studied in this paper.On the basis of full convolution neural network,a scale pyramid space is cascaded to form a multi-scale corner detector.The scale pyramid space detects the key points of the target to be segmented from different scale feature maps,and learns the deeper context feature information of the image,which improves the accuracy of semantic segmentation.Aiming at the difficulty of manual labeling data sets,a weak supervised semantic information detector is designed.The target frame annotation feature information and image label feature information are learned in the training stage of the neural network.After several iterations,more semantic label information can be learned in the weak supervised semantic environment,and the weak supervised semantic segmentation is realized.The experimental results show that the proposed algorithms can effectively solve the related problems in the process of semantics segmentation and improve the accuracy of semantic understanding of traffic scenes.