Research On Probabilistic Vehicle Localization In Complex Scenarios

Autonomous driving can improve traffic efficiency,which helps strengthen transportation network.Vehicle localization,to determine the ego vehicle's position,is the essential technique of autonomous driving.The main localization approach is to observe the surrounding using on-board sensors and then match measurements with prior maps.However,traditional methods may have low accuracy or even fail when the vehicle drives in real complex scenarios,such as urban canyon,vehicle state changing,illumination variation,and traffic congestion.This thesis aims to solve the problem of reliable vehicle localization in complex scenarios using probabilistic theory,and proposes approaches based on features of the standardized trajectory,terrain and map-based road information.The detailed research works are as follows:Trajectory-based localization approaches can estimate the ego vehicle position efficiently in GPS denied environments.However,the reliability of existing approaches decreases when the ego vehicle changes the lane.To address this problem,this thesis proposes a probabilistic localization method based on the standardized trajectory feature.The vehicle trajectory is standardized to eliminate the influence of lane changing.Then the bayesian-filter-based localization is developed.To improve the convergence speed,the proposed method introduces the terrain feature to get a coarse localization result by linear prediction.Experimental results demonstrate the proposed method has higher accuracy and speed than traditional trajectorybased methods especially in lane changing scenarios.Terrain-based localization approaches can determine the ego vehicle position efficiently when the road shape is simple.However,the error of terrain measurements increases during braking,leading to wrong position estimation.To address this problem,this thesis proposes an adaptive terrainbased localization method.First,the relation between the noise of models and the vehicle acceleration is studied statistically.Then,the state transition model is developed considering the acceleration.Finally,the localization algorithm is proposed based on adaptive particle filters.Experimental results demonstrate that compared with traditional methods,the proposed method can localize the vehicle efficiently and accurately,even when the observation changes dramatically.Road-feature-based localization approaches can perform lane-level positioning.However,the error of measurements increases during illumination variation or occlusion.To address this problem,this thesis proposes a map-supervised-feature-based probabilistic localization method.First,the probabilistic lateral road feature is designed to express the distribution of ground signs.Then,the CNN-based perceptual module is developed to reduce the effects of illumination variation and occlusion,which takes raw images from around view cameras as inputs.Aiming at the difficulty of obtaining datasets,this paper proposes a training method based on map supervision without special manual labeling.Finally,in order to reduce the positioning noise caused by multi-modal single-frame errors,the method exploits probabilistic data association to implement recursive positioning.The experimental results demonstrate that the proposed algorithm can achieve sub-meter level accuracy in complex scenes such as dark nights,multiple vehicle occlusions,and lane number changes.