Research On Localization In High-speed Railway Station Based On Image Retrieval Technology And Hierarchical SFM Method

With the continuous expansion of the scale of high-speed railway stations,the structural layout of the station has become increasingly complex.There are many functional areas in the station,and the space is large,resulting in the increasing demand of passengers for positioning in the station.Compared with traditional outdoor positioning,the high-speed railway station scene has the characteristics of complex environment,high building similarity and large signal transmission loss,which makes the positioning complexity in the high-speed railway station more complicated.When it comes to personal issues such as passengers entering and leaving the high-speed railway station in time,the accuracy of scene recognition and positioning in high-speed railway stations is often relatively high.At present,many indoor positioning technologies are located by setting up signal equipment on site for signal recognitinon,but this method will inevitably be disturbed in high-speed railway stations with high crowd density,which will cause great trouble to the identification on site,and equipment layout and maintenance costs also need to be considered.Therefore,this paper starts from the recognition of the scene in the high-speed railway station,and uses the three-dimensional scene reconstruction technology to restore the target scene in the high-speed railway station,so as to achieve relatively accurate positioning in the station.An indoor positioning method based on image retrieval technology and threedimensional reconstruction technology is proposed in this paper,and the key algorithms for scene restoration and reconstruction and scene feature recognition are studied.(1)In terms of scene restoration and reconstruction,SFM(Motion Structure Restoration)is used to reconstruct the sparse point cloud of the scene.The basic process is analyzed and designed,and the related key technologies are studied.Different reconstruction processes and ideas are classified into three categories: incremental,global and hybrid.Through the designed experiments,the three traditional methods are compared and evaluated in terms of rotation error,translation error and the number of recovered 3D points.Its features,advantages and disadvantages are summarized.(2)Based on the comparative analysis of traditional SFM methods,the hybrid SFM method is improved,and the details are adjusted with reference to the idea of using the advantages of incremental and global methods.A hierarchical SFM method based on bag-of-words tree.In the global process design,the feature vector obtained from the image is extracted by the SIFT method,and the image data set is converted into a tree structure by using the bag-of-words tree principle,so that the image set is processed.Classification,remove branches with large rotation errors,and complete the global reconstruction process.The high-speed railway station image set taken by the camera is used for experiments,and the improved method is compared with the traditional hybrid method to verify its effectiveness.(3)In the aspect of scene recognition,this paper proposes to use the convolutional neural network as the basis,and assist in the recognition,matching and retrieval of image features by means of the nearest neighbor search algorithm.According to the requirements of scale invariance and stability required for scene feature recognition,the Dense Net121 model is selected as the basic model,and the locality-sensitive hash algorithm and KD tree algorithm in the neighbor search algorithm are compared and experimentally analyzed.The characteristic image retrieval stage uses a method combining Dense Net121 model and locality-sensitive hash function,and the correct rate and effect of this method are verified.In the test phase,take the relevant facilities and signs in Hengshui North Station as the experimental scene,use the camera to shoot the corresponding scene in the shooting area of the specified distance to obtain the image data set information,and test the positioning accuracy,and the results show the coordinates of the field measurement The abscissa error with the calculated coordinates is between 0.020% and 0.015%,and the ordinate error is between 0.00172% and 0.00164%.The actual positioning distance can be controlled within a distance of 3 to 4 meters,which proves that the indoor positioning method can Effectively complete indoor relative accurate positioning.