Research On Real-Time Semantic Segmentation Of Outdoor Scene Based On Deep Learning

In recent years,with the continuous development of artificial intelligence and deep learning technology,a variety of outdoor intelligent products emerge in endlessly,such as unmanned driving,unmanned aerial vehicles,smart city transportation and other aspects have broad development prospects.Among them,the semantic segmentation technology based on deep learning plays a pivotal role in visual perception.It can identify and classify objects at the pixel level and restore the shape of objects.It is very suitable for those objects that do not use block diagram identification,such as streets,sidewalks,houses,lamp posts and other objects.At the same time,compared with the traditional method,it has higher accuracy and better generalization.However,there is still a lot of room for improvement in the processing speed of high-precision network.Although the level of hardware is constantly developing and improving,the design and optimization of network is still our goal to achieve.In order to ensure the smooth and normal use of the network on outdoor equipment,it is necessary to design an efficient real-time network.Firstly,this thesis introduces the research background of the subject,briefly describes the research status and development trend of semantic segmentation technology based on deep learning in various kinds of artificial intelligence applications at home and abroad,and briefly explains the research content and structure of this thesis.Next,this thesis introduces the overall architecture of the common semantic segmentation network and modules,the basic architecture of each part,common semantic segmentation network is difficult to balance calculation speed and accuracy,so the depthwise separable convolution is used to replace traditional convolution,collocating 1×1 traditional convolution with expansion and compression channel,coupled with skip architecture,in order to achieve calculation speed and accuracy of the trade-offs.At the same time,when designing the loss function,considering the large gap between categories,the loss function needs to be weighted to balance the gap between classes.The network can achieve 14.0 FPS on NVIDIA Jetson TX2 board hardware and 62%mIoU(mean intersection-over-union)on Cityscapes datasets.For lane line segmentation,another branch is added on the original real-time segmentation network to obtain a four-dimensional vector for each pixel.Finally,clustering algorithm is used to cluster different lane line pixels.The network can achieve about 50fps on NVIDIA Tesla P100.The results show that the two network can achieve real-time effect and have good accuracy.