Highly Reliable Lane-level Fusion Positioning Methodology For Intelligent Vehicles In Complex Traffic Environment

In order to effectively reduce driver errors,improve traffic safety and enhance efficiency,intelligent vehicles with advanced driver assistance functions have attracted great attention in recent years.In fact,typical intelligent assisted driving behaviors involve lane-level positioning and judgment.Therefore,how to use low-cost means to achieve highly-reliable lane-level positioning has become one of the bottlenecks for the in-depth development of intelligent vehicles.Existing works mainly introduce new auxiliary sensor methods or propose fusion algorithms with the ability of prediction and compensation based on the integration of satellite/inertial navigation system(INS).However,they still have obvious limitations.The former has poor environmental adaptability,while the latter has insufficient prediction generalization and correction performance.In order to overcome such limitations,inspired by the new development of deep learning,this dissertation investigates the highly reliable lane-level fusion positioning methodology for civil intelligent vehicle in complex traffic environment from the levels of auxiliary sensor method and systematic fusion algorithm.Firstly,taking simulating human eyes’ lateral positioning as a new idea,this dissertation focuses on the key parameter estimation methods with strong environmental adaptability,which can provide effective lateral position and yaw angle observation information for global fusion positioning.Then,an intelligent fusion algorithm with superior error prediction and compensation ability is studied to further improve the overall performance of vehicle positioning.The main contribution and innovation of this dissertation can be summarized as:1)A lateral position estimation method based on unstructured feature is proposed to cope with the problem that existing approaches highly depend on structured traffic facilities or signs,which integrates two deep neural networks(DNNs)into one framework to simulate the lateral positioning mechanism of human eyes.Firstly,a convolutional neural network(CNN)-based road detection network is built to obtain the road area and then it is regarded as a highly reliable reference object for lateral position estimation for the first time,breaking through the limitations of traditional positioning references.Then,the hidden lateral position information is mined from the road image to construct a parameter estimation model based on multi-layer perceptron(MLP).The proposed method solves the problem of poor environmental adaptability to lateral position estimation,obtaining accurate and reliable estimation results in unstructured road scenarios with low traffic density.As far as the author’s knowledge is concerned,there are few reports on the using road area for vehicle lateral positioning.2)Inspired by the human-like visual lateral positioning model,a lateral position estimation method based on multi-level robust feature fusion is constructed to deal with the problem that existing approaches cannot effectively tackle the partial occlusion of road surface in medium traffic density.Firstly,an attention-guided network is designed for road detection to ensure that it can still extract accurate road areas even in partial occlusion situations,so as to achieve the occlusion-handling ability at the object detection level.Besides,a fully convolutional denoising autoencoder with lateral connections is built to extract robust localization features from the road areas.With the help of a novel denoising mechanism,the network improves its robustness to occluding objects at the feature extraction level.Finally,a long short-term memory network is established to learn the long-term dependencies of localization features and memorize the similarity of lateral positions under different degrees of road occlusion,which can reduce the estimation difficulty of the network when the occlusion degree increases and enhance its robustness at the decision output level.By fusing these multi-level robust feature learning mechanisms,the proposed method can achieve accurate and reliable lateral positioning in partial occlusion scenarios.3)A lateral position and yaw angle estimation method based on road area reconstruction is proposed to solves the problems that existing lateral position estimation approaches are difficult to adapt to severe road occlusion scenes in high traffic density and the reliability of yaw angle estimation is poor.This method consists of three modules,which are integrated into a framework in a human-like manner to achieve a high degree of modeling human-eyes lateral positioning.Firstly,a road area detection and occluding object segmentation module is constructed to obtain their shape information based on siamesed convolution network.Furthermore,a road area reconstruction module is designed to automatically infer the occluded road area and obtain the complete one based on conditional generative adversarial network,in which the features extracted from the shape information acting as the basis and guidance,respectively.Finally,a lateral position and yaw angle estimation module is built to extract key spatio-temporal features from continuous complete road-area images and make the prediction based on 3D convolution and long short short-term memory network.Benefitting the automatic image inpainting mechanism,the proposed approach exhibits strong robustness to severe road occlusion scenarios in high traffic density.To the best of our knowledge,the model simultaneously overcomes the challenges of severe road occlusion and poor reliability for localization reference objects,which is the first time in the field of vehicle lateral positioning.4)According to the availability of GPS,a direct fusion positioning scheme based on observation information enhancement and an intelligent fusion positioning scheme based on error prediction and compensation are proposed respectively to solve the problem that traditional GPS/MEMS-INS integrated positioning scheme is difficult to achieve highlyreliable lane-level positioning in complex traffic environment.When GPS is available,an extended Kalman filter(EKF)is used to fuse the GPS pseudo range(subjective observation),lateral position and yaw angle(auxiliary observation),which can comprehensively improve the vehicle positioning accuracy by introducting high-quality observation information.When GPS fails,a convolutional denoising autoencoder-based MEMS-IMU data denoising model,a multi-observation differential fusion strategy and a long short-term memory network-based position error prediction model are constructed to address the problems of weak generalization ability and limited compensation effect for existing error prediction and compensation approaches.Compared with the methods that directly model the noisy MEMSIMU data,the intelligent fusion positioning scheme can learn the changing characteristics of position errors more accurately by modeling robust latent features,and further improve the positioning performance of vehicles in complex traffic environment.