Research On Visual Indoor Positioning Based On Image Features

With the rapid development of urbanization,people spend more than 80% of their time in indoor spaces,so there is a strong demand for fast and accurate indoor positioning.Indoor positioning technology based on visual sensors has become a research hotspot due to its low cost,convenience,practicality,and easy deployment.At the same time,with the gradual maturity of computer vision theory,the visual indoor positioning method based on camera and image has high research value and broad application prospects.This paper uses the Kinect V2 camera as a carrier and the feature information in the camera image,combines the feature points of the image and its depth information to construct a visual indoor positioning technology framework based on image features,which is composed of two parts: offline database construction and online positioning.At the same time,this paper relies on experiments in real indoor scenes to verify the feasibility of the method,and combines traditional surveying and mapping techniques to verify and evaluate the indoor positioning accuracy.A set of visual indoor positioning solutions proposed in this paper provides a useful reference for the research and application of indoor positioning methods.First of all,this paper studies the geometric theory of the camera from the camera sensor model,imaging principle,coordinate system conversion relationship and camera parameter solution.The parameters of the Kinect V2 camera are obtained by Zhang Zhengyou calibration method and the reprojection error is calculated.Then,this paper performs registration fusion on the RGB image and the depth image acquired by the Kinect V2 camera to obtain the 3D coordinate mapping of the image pixels.At the same time,an image collection cloud platform was independently built to perform point-by-point interval sampling in two different indoor scenes.The ORB features are extracted from the collected images,and the feature clustering is realized by the K-means algorithm.Then introduce the k-d tree dictionary to classify,store and express the scene images.Thus,an indoor image database containing image feature information and geographic location information is constructed.Next,this paper combines the construction process of the visual bag of words model,introduces the TF-IDF weighted model to achieve feature indexing,and then uses the similarity calculation theory to retrieve matching image pairs in the database that are similar to the user's query image.In addition,the improved pose estimation problem based on the principle of perspective n-point projection is applied to indoor positioning.The EPn P algorithm is used to complete the 3D-2D feature point matching camera matrix solution and pose solution,and finally realize the indoor online positioning process.Finally,this paper uses a total station to compare and verify the reliability of the pole-constraint method and the EPn P algorithm,and calculates the error and evaluates the accuracy of the positioning results in an indoor and a corridor through the form of measured coordinates.The experimental results show that if the feature texture in the indoor environment is richer,the matching result of the image pair will be more accurate and the accuracy of visual indoor positioning will be higher.In addition,the EPn P algorithm based on3D-2D feature point matching in the same indoor scene has higher positioning accuracy than the pole-constraint method,increasing by 15.2% and 19.3% respectively.The positioning error is kept within 15 cm,which basically meets the requirements of accurate indoor positioning.