Research On Key Technologies In Vision-based Real-time Light Rail Localization System With High Accuracy

As an important component of urban transport system,light rail system relieves the city traffic pressure,and becomes an emerging public transport gradually in recent years.At present,urban light rails are mostly operated by manual control and share the roads with cars and buses.Intelligent rail-assisted driving system is the development direction of future light rail transit system.Light-rail localization is an important part of the light-rail driving assistance system.It is of great significance to study highprecision real-time localization of light rail.As an emerging localization method,vision-based localization methods have the advantages of no special sensors,low maintenance costs,and non-susceptible to external signal interference,etc,which have been widely used in localization and navigation systems for cars and robots.Its principle is to convert the visual information captured by the camera into the map information,achieve the optimal matching of the current visual information and historical reference information through the scene matching method,and then obtain the current position information.However,there are still many problems that need to be solved in the current vision-based light rail localization.By considering practical needs like real-time,high-precision,low maintenance costs and scalability,the dissertation focused on key technologies in vision-based real-time light rail localization system with high accuracy.The main research contents are as follows.(1)An orientation and scale invariant binary feature based on Haar wavelwet is proposed.To solve the problem of low insufficient robustness insufficient robustness of intensity differnece/local gradient based binary features,such as rotation,illumination,blurring,and scaling changes,the proposed method first calculates the Haar wavelet response in the feature description region.A binarization sampling pattern is generated to extract the local binarization features of the image.The description region is normalized by the orientation of the centroid vector.The multi-scale Haar wavelet operator is used to calculate the feature vectors at different scales to ensure the orientation and scale invariance of the image features.The experimental results show that the binary descriptors extracted by this algorithm are superior to other feature description algorithms in feature point matching,significantly improving the distinguishability of the local binarized description of images.It has a high degree of matching accuracy and speed when blurring,degrading,view-point and illumination changes occur in the images.(2)An unsupervised scene feature area detection and key frame extraction method based on discrimination is proposed.For the light rail video sequence scene,there are a lot of problems from useless areas,lighting,viewing angles,occlusion and other interferences.This method designs a discrimination score to measure the saliency of pixels in the scene and uses a relative threshold algorithm to extract the most representative feature area of the current scene in the high frame rate video,which greatly reduce the amount of scene feature extraction and matching in online scene matching.In order to solve the problem of different road segments but similar scene content,the key frames of the reference sequence are extracted based on the regional discrimination scores to provide a suitable search window for scene tracking,so as to avoid large tracking localization errors caused by similar scenes.Experimental results show that the method can significantly improve the accuracy of scene matching while significantly shortening the time of scene matching procedure.(3)A fast scene feature extraction method based on feature discrimination score and correlation coefficient is proposed.Aiming at the problem of high degree of similarity and indistinguishability between continuous scenes in high-frame-rate light rail video sequences,the method establishes pixel pair sets within the image feature area and measures all pixel pairs of in the set of pixel pairs by defining a calculation method of discrimination scores.In order to reduce the redundant information in the generated binary descriptors,based on the saliency scores,a greedy algorithm is used to filter out pairs of pixels in the pixel pair set with low correlation,so as to increase the description power of scene descriptor.Experimental results show that scene feature descriptors extracted based on this method can distinguish scene images with extremely high similarity in high frame rate video and achieve high-precision real-time localization of light rail vehicles.(4)A high-precision light rail real-time localization method based on globallocal scene features and key frame retrieval is proposed.For the disturbance caused by scenes with similar appearances,such as stations,in the light rail video sequence,and the high similarity of continuous scenes at high frame rates,the proposed method realizes high-precision real-time localization for light rail,by combining the feature area detection,key frame extraction,and scene feature extraction methods proposed above,with a key frame retrieval strategy,global-local scene features and SeqSLAM scene matching method.A high-precision light rail real-time localization system based on visual information was implemented.Experiment results show that the system can achieve accurate localization of the light rail in a high frame rate,and can maintain high matching accuracy in extreme situations such as large-area scene occlusion.While meeting the high-precision needs of the light rail system,it significantly improves the real-time performance.