Accurate Vision-based Localization For Intelligent Vehicles

In recent years,with the rapid development of computer technology,information technology,automation technology and vehicle manufacture,People's transport demand begin to be varied,and the vehicle is becoming more and more high quality requirements,thus,intelligent vehicle technology begins to be widespread concern in the community,especially,in academia and industry,setting off a boom in research and development.And the precise Localization of the vehicle is one of the most key technology of intelligent vehicle.Generally,traditional localization methods use IMU as additional sensor to help GPS increase the localization accuracy.However,the high accuracy IMU and vision SLAM have limitations such as IMU suffers from high cost and V-SLAM suffers from accumulative errors.At the same time,the urban environment is more and more structured,and the road scene level is more and more distinct;making it possible for camera image to extract rich and effective features.Hence,this paper is focus on high accuracy vehicle localization by using low cost GPS and vision sensors.This paper has studied for artificial landmarks included and natural scenes,respectively.In artificial landmarks study,Road Sign Inventory(RSI)database is established,more specifically,the database contains high accuracy road sign position data and attribute information of road sign,and their holistic features by ORB and local feature by BOW are extracted.In natural scenes study,high accuracy road scene model is established,more specifically,the model contains precise GPS trajectories,3D point cloud and the corresponding 2D features,and holistic features by ORB and SURF.In next step,multi-scale localization is accomplished,which contains normal GPS localization,image-level localization and metric localization.In image-level localization,H-KNN method is proposed to image identification.Experimental results show that the localization errors in two different road scenes are 15 cm and 12.8 cm,respectively,and the author carried out a simple fusion processing of the two localization methods.Contributions and study contents are summarized as follow:(1)Proposed road sign recognition algorithm by using ORB and BOW feature.This paper first proposes a traffic sign holistic feature descriptor based on ORB patch.Meanwhile,local features of traffic sign are modeled by BOW.Vision representation is generated by holistic and local features.Finally,traffic signs are recognized by fusing H-KNN and traffic sign representation.The proposed H-KNN model is a generalized model that can be used to fuse different spatial data to form the final voting space.(2)Established a high accuracy road scene model.Road scene model contain 3 components: high accuracy trajectories,3D data,2D holistic feature and local feature.Trajectories are collected by IMU/GPS,and corrected by RTK,holistic image features are described by ORB and SURF,respectively.Local features are extracted by ORB descriptors.3D data are obtained by stereo camera using 3D reconstruction.(3)Proposed a high accuracy vehicle localization method by using artificial landmark.RSI is first establishment,which is mainly for traffic signs to collect its location and its associated attribute information,including color,shape,orientation and so on.In vehicle localization,multi-scale localization is used,which contains GPS coarse localization,image-level localization and metric localization.(4)Proposed a vehicle localization method based on natural scenes.First of all,road scene model is established.The performance of the typical feature point extraction operator is compared by big samples,feature operators are SIFT,SURF and ORB,respectively.In Localization step,the multi-scale localization algorithm is applied,and vehicle localization is completed based on natural scenes.And the author evaluated the effectiveness of the algorithm using different ORB feature number extracted from the images.Especially in the process of image scale localization,this paper uses ORB and SURF holistic feature to recognize successive scenes image.