Research On Semantic Segmentation Method Of Robot Outdoor Perception Image Based On Deep Learning

With the rapid development of deep learning and the widespread use of mobile robots,autonomous driving has gradually become a research hotspot.Environment perception is the most important part of automatic driving technology,and the purpose of environment perception is to distinguish the content of the environment.Therefore,accurate and efficient image semantic segmentation method is becoming more and more important.This paper will study the semantic segmentation method of robot outdoor perception image based on deep learning,propose a semantic segmentation network model that can speed up the network operation speed while ensuring the accuracy,and a high-precision semantic segmentation network model that can enhance the boundary information.The main research contents of this paper are as follows:Firstly,the important components of convolutional neural network and the principle of each specific operation are studied.Several representative image semantic segmentation network models are analyzed.The image preprocessing method and image evaluation criterion are designed.Secondly,a full convolution image semantic segmentation network based on multi-scale feature fusion is designed.An improved multi-scale module and an improved residual module are proposed and introduced into the network model in this paper.The improved multi-scale module is mainly composed of ordinary convolution and dilated convolution,which makes full use of the advantage of dilated convolution to obtain the image feature information.By comparing and analyzing the advantages and disadvantages of batch normalization and group normalization,group normalization was creatively applied to the residual module,and Leaky Re LU(LRe LU)activation function,which could improve the robustness of the network model,was applied to the residual module.In order to make the reference information of the network prediction more comprehensive,this paper adds an image preprocessing module to the network model.By analyzing the advantages and disadvantages of convolution operation and pooling operation,and taking advantage of convolution,a full convolution network model is designed,and comparative experiments are carried out.The experimental results show that the full convolution network model can improve the accuracy of predictive segmentation.Thirdly,an encoding-decoding network is designed to enhance boundary information.This paper divides the network model into two parts: encoder and decoder.By analyzing the advantages and disadvantages of bilinear interpolation and deconvolution,two kinds of network models using deconvolution for up-sampling operation are designed,and comparative experiments are carried out.The experimental results show that the fusion of the characteristic information of four and eight times down-sampling and corresponding characteristic information is more effective.The residual chain structure is innovatively introduced,and the shallow features extracted from the network are firstly convolved and then fused with the deep abstract features,so as to reduce the difference of the fusion part and improve the accuracy of the algorithm.The loss function of the network was optimized,that is,a kind of Focal Loss function that could alleviate the non-equilibrium of the sample was selected,and the experiment was conducted to compare it with the traditional cross entropy loss function.The experimental results proved that Focal Loss function was more conducive to the training of the model.Finally,a traveler IV mobile robot in the laboratory was used to collect images of campus street scenes and make their own data sets.The network model designed in this paper is compared with the produced data set and analyzed for different application scenarios.The experimental results prove that the multi-scale feature fusion full convolutional network model is more suitable for predictive segmentation in simple scenarios,and enhances boundary information encoding-decoding network model is more suitable for predictive segmentation in complex scenes.