Research On Outdoor Scene Semantic Segmentation Algorithm Based On Deep Learning For Intelligent Vehicle

In recent years,the problems of frequent occurrence of security accidents,urban traffic jams those are presented by traditional vehicles serve as the impetus to drive the rapid development of technologies for driverless and the market applications.As the key technology for the environmental awareness in the driverless system,outdoor scene semantic segmentation algorithm based on deep learning has been held in high regard by the relevant researchers.The main research contents of the thesis are focusing on the outdoor scene semantic segmentation algorithm based on deep learning.The specific research contents of the thesis are as follows:1.Analyzes the strengths and weaknesses of outdoor scene data in the most used semantic segmentation data set(Cityscapes,Mappilary,Camvid ect.).The label tags of the collected road scene data were manually marked using the Label Me Tool,determine the semantic segmentation data set used for the research of the semantic segmentation algorithm in this thesis and evaluation indices those are appropriate for the research,and proposes the evaluation function to assess the average segmentation prediction accuracy of the model.2.Builds the semantic segmentation model that optimizes the analysis of the global context feature.Based on the Deeplab network,the model combines with spatial pyramid pooling feature aggregation module to better analyze the global context feature in the model.Compared to the Deeplab model,the sample data from the semantic segmentation data set we build for the thesis test shows that the average segmentation prediction accuracy of the model we build is 0.27% higher than the original one.The results of segmentation prediction accuracy of large object categories(such as road)and traffic sign in the scene significantly improve(4.74%).3.To tackle the problems of the low computational efficiency and the slow segmentation inference speed of the semantic segmentation model that optimizes the analysis of the global context feature we build in the thesis,this thesis builds the multi_size feature concatenate semantic segmentation model.The model combines Shuffle Net model and skip connection architecture.The sample data from the semantic segmentation data set we build for the thesis test shows that the average segmentation prediction accuracy of the multi_size feature concatenate semantic segmentation model is 0.5928.The segmentation inference speed of the model is around five times as fast as the one of the semantic segmentation model that optimizes the analysis of the global context feature we build in the thesis,the specific value is 22.17 FPS which nearly meets the requirement for the real-time understanding.The multi_size feature concatenate semantic segmentation model we build in this thesis is suitable for the semantic segmentation understanding of the simple urban road scenes for the intelligent vehicle.4.Studies the method to improve the average segmentation prediction accuracy of the semantic segmentation model which is based on deep learning from the network architecture.In this thesis,we set the semantic segmentation model that optimizes the analysis of the global context feature as the experimental model.Focusing on the bilinear interpolation algorithm adopted in the model,we propose the special dense upsampling convolution module(DUC module)that is unique for the experiment model to improve the average segmentation prediction accuracy of the experimental model.The result of the experiment shows that the model optimization module we proposed recovers the subtle context features which are missed in the original semantic segmentation model caused by using the bilinear interpolation algorithm to implement upsampling operation to generate the model prediction output tensor.Compared to the original semantic segmentation model,the average segmentation prediction accuracy of the model is increased by 3.92%.