Global Localization Method Based On Semantic Segmentation For Intelligent Vehicles

In recent years,the intelligent driving system has drawn more and more attention from academia and industry.Global localization is essential for intelligent vehicles when navigating on the urban road.Visual localization methods are widely applied because of rich features and low price.Generally,visual localization methods based on visual features are not robust to occlusion and light changes.This will lead to mismatch easily,resulting in a decline in accuracy.Therefore,this paper proposes a method of global localization using ground marks with semantic information such as lanes,stop lines,zebra crossing and arrows as landmarks.Traditional rule-based ground marks detection methods are not adaptable,so this paper proposes to use the semantic segmentation method based on deep learning to extract the ground marks.The contour features of ground marks are used for localization.And the features are fused with high-precision GPS data to create a priori map,the features extracted online are matched with the map to estimate the global localization using the ICP(iterative closest point)algorithm.Finally,the global localization is fused with other sensors data by the EKF(Extended-Kalman Filter)for better accuracy.Experiment results show that this method is robust,and can reach decimeter-level accuracy,which can meet the demand for intelligent vehicles.However,this method requires lane-level lateral localization as initial value.To solve this problem,we propose a method to extract road boundaries by using semantic segmentation based on depth learning.Then the road boundaries are matched to the map to acquire lane-level lateral localization as initial.The experimental results show that this method can provide lateral initial location for global localization,at the same time it improves the robustness to occlusion.