Research On Road Semantic Segmentation Method In Front Of The Vehicles Based On Global Feature Transformation

Free space detection in front of vehicles is a crucial basis for the environment understanding in driver assistance systems.Its related road semantic segmentation is a research highlight in computer vision and image processing.Recently,deep convolutional neural networks have been implemented to extract high-discriminable features from road images accomplishing binary road predictions.However,the mainstream road segmentation methods suffer from two challenges: one is the hollows and interruptions in the results of the convolution networks’ predictions because of the texture variation in the condition of light changes in the complex road environment;the other is the problems of mismatched and excessive road boundaries’ segmentation causing by ambiguous road edges.Therefore,the thesis focuses on the research on road semantic segmentation method based on global feature transformation in front of the vehicles to address the above problems.This thesis develops the research in the following three aspects:1.Most of the current algorithms suffer from the problem of voids and interruptions in road segmentation results.This thesis proposes a road semantic segmentation method based on hierarchical global feature transformation.Firstly,symmetric dual-branch feature encoders are employed to extract local features from RGB and depth images.Next,hierarchical attention consolidation units are designed to model global features parallelly under the global receptive field.A cross-attention transformation unit is constructed to model the similarity of RGB and depth features to form an effective feature fusion.By capturing road features under the global perceptual field through hierarchical global feature transformation,the problem of voids in road segmentation results is effectively alleviated.This method improves the original network F-score and recall by2.49% and 3.08%,respectively.2.In response to the fact that previous methods have mismatched and excessive road boundaries,this thesis proposes a road semantic segmentation network based on progressive uncertainty analysis to explicitly model road edges.Firstly,the progressive uncertainty analysis module is designed to refine the edges between roads and non-roads from coarse to fine.Secondly,upgraded uncertainty loss is designed to remove misleading evidence in the Dirichlet distribution based on road probability concentrations to constrain the correct road area prediction.In addition,a multi-scale global-local attention interaction module is designed to fuse multi-scale features via global feature transformations,to generate deep semantic features with more robust feature feedback.The method models road edges explicitly while maintaining efficient computation,predicting the road confidence map to improve the accuracy of road boundaries segmentation effectively.3.Since the attention drift and semantic misalignment due to unlearnable position encoding in global feature transformation,this thesis proposes a depth feature position encoding-guided road semantic segmentation framework.Firstly,a pyramid Transformer backbone is constructed to perform feature extraction under the global perceptual field.Meanwhile,the position encoding is designed using the locality of the depth features extracted by the convolutional neural network to model the real spatial relationship of pixels.Finally,a multi-level global feature transformation optimization module is designed to extract local features in a windowed manner to enhance detailed understanding.The method aligns target details in global understanding to optimize road segmentation,ranking first on the KITTI dataset with a 97.53% F-score.In summary,this thesis proposes an innovative road segmentation method based on hierarchical global feature transformation,progressive uncertainty analysis,and depth position encoding guidance to address the two challenges in existing road segmentation methods.The algorithm proposed in this thesis is validated in detail on two public datasets of traffic environments,KITTI and Cityscapes,to demonstrate that it significantly improves the completeness and edge fineness of road segmentation and achieves highly accurate and efficient road semantic segmentation.