Research On Semantic Segmentation Of Road Scene Image In The Field Of Driverless

With the continuous development of driverless technology,the perception demand for road scene is also increasing,the key is computer vision technology.In the field of driverless,the typical application of computer vision is the image semantic segmentation for road scene.This thesis is based on Deeplab V3 + network and fully connected conditional random field,aiming at the problems that may appear in image semantic segmentation,such as the loss of details of pixel feature information,the loss of spatial context feature information,the blur and discontinuity of edge pixels,this thesis uses the main and sub network mode of image semantic segmentation,and combines with a variety of improved algorithms to carry out the following three aspects of research work :Firstly,for the problem of detail loss of pixel feature information and spatial context information loss in the process of convolution and pooling operation of full convolution network,in Deeplab V3+ network’s coding end,the dense hole convolution method based on coprime factor is used to replace the standard hole convolution,which can expand the receptive field of convolution kernel,extract and save image feature information efficiently;then in the pyramid pooling structure,the input feature of low hole rate sampling layer is fused with the output of this layer,and the result is used as the input feature of subsequent high hole rate sampling layer,at the same time,the output features are pooled by global average to obtain the spatial feature information of multi-scale pixels.In the decoding end of Deeplab V3+ network,the shallow feature information of the decoding network is processed by cross layer fusion,and the spatial context information of the pixels is further saved,so as to build the main network of image semantic segmentation Deep Pnet.And on the road scene dataset: cityscapes,the comparison test of Deeplab V3+ and Deep Pnet is carried out.The experimental results show that Deep Pnet network can not only finish the semantic segmentation of road scene,but also effectively process the detail information and spatial information of pixels,which improves the accuracy of image semantic segmentation.The experimental results prove the rationality and feasibility of the improved method.Secondly,for the problem of fuzzy and discontinuous edge in image semantic segmentation,based on fully connected conditional random field,g Pb edge detection module is introduced into fully connected conditional random field model to realize the extraction and preservation of edge pixel feature information Then,a robust high-order potential energy term is introduced to enhance the processing ability of the fully connected conditional random field model for edge high-dimensional information,so as to build the sub network of image semantic segmentation g Pb-HOFCCRF.The main network of image semantic segmentation Deep Pnet is integrated with g Pb-HOFCCRF model based on fully connected conditional random field model to construct a complete image semantic segmentation network Dnet-CRF for road scenes.Finally,on the road scene dataset cityscapes,the Deeplab V3+ and Dnet-CRF network are tested.The experimental results show that the improved algorithm not only overcomes the fuzzy and discontinuous problems of edge pixel segmentation,but also further improves the accuracy of image semantic segmentation.The experimental results prove the feasibility and superiority of the improved method.