Research And Implementation Of Semantic Segmentation Of Urban Street View Images

Environmental perception technology based on image semantic segmentation can help autonomous driving system understand the surrounding driving environment and provide data support for the subsequent decision and control of the autonomous driving system.In this paper the semantic segmentation algorithm for urban streetscape images is studied,the main research content and innovation work are as follows:First,a semantic segmentation dataset of domestic urban streetscape images is constructed.The existing semantic segmentation dataset of urban streetscape images lacks ground traffic sign targets such as lane lines and straight arrows,which are crucial for the safe driving of autonomous vehicles,and these datasets are mainly constructed based on foreign urban streetscapes.To address this problem,obtain domestic urban streetscape images through the car recorder,then use Labelme software to manually label the set targets and generate VOC format dataset,and finally enhance the dataset by geometric transformation and color adjustment to complete the construction of the dataset.Secondly,a lightweight semantic segmentation model for urban street images with improved Deep Lab V3+ is proposed.To address the problem that the Deep Lab V3+semantic segmentation model is too large in number of parameters and too slow to be applied to real time interactive autonomous driving scenarios,the Xception network is replaced by the lightweight Mobile Net V2 network as backbone of the model in the coding area,and the depthwise separable convolution is applied to the ASPP module.At the same time,a dilated convolution branch is added to the ASPP module to improve its ability to extract low-resolution targets.To address the problem that the shallow features extracted by the backbone network are not fully utilized during the upsampling of the deep feature map in the decoding region,a multiscale feature fusion network is proposed,in which three shallow features at different scales are fused during the upsampling of the deep feature map.And the spatial attention module and channel attention are used to optimize the shallow features and deep features respectively before feature fusion.Experiments show that the improved Deep Lab V3+ semantic segmentation model in this paper has improved in both segmentation accuracy and speed.Finally,a semantic segmentation software platform for urban streetscape images is designed and implemented based on Py Qt5.The platform supports users to invoke Deep Lab V3+ semantic segmentation model to semantically segment the city streetscape images in local disk by mouse click,and can save the processed images to the local disk.