Visual Indoor Localization Based On Real Images

Indoor visual positioning based on real images realizes indoor positioning by identifying a real-world indoor image and the specific location of the image in the building,which has the advantages of entirely independent of any infrastructure and a high degree of visualization.However,due to the fact that the real images taken in practical applications usually have problems such as unbalanced exposure,jitter,and occlusion,and the high cost of making image data sets,this scheme faces a serious challenge to the practical application.Based on the vision positioning idea of region division,a low-cost and efficient data set acquisition and calibration method is proposed,and a convenient indoor vision positioning system based on real scene images is designed and implemented by using the deep neural network to identify images for position reasoning.The production of the data set is divided into two parts: original video preprocessing and video sequence calibration: firstly,the original video is preprocessed.Aiming at the jitter problem of handheld shooting,a secondary filtering algorithm based on sliding window is proposed by analyzing the gradient change between video frames.For the preprocessed video sequence,sparse optical flow method is used to track feature points and solve the essential matrix of camera based on the principle of epipolar geometry.Then,the singular value decomposition is used to restore the video shooting path to calibrate the image spatial coordinates.In order to adapt to the limitation of mobile computing resources,lightweight convolutional neural network Moblie Net V3 is adopted as the core network to identify image.And to solve the problem of the inconsistency of distribution of the data set,locating method based on region partition is put forward,which means that all the images belonging to same region are quantified with same label in the training stage of the model.And in the model reasoning stage,the real scene image is used as input,and then the neural network is used to classify it to the corresponding sub-region,and the localization problem is converted to the classification problem.Taking the fourth floor of the South first Building,Huazhong University of science and technology as the experimental site,the comparative experiments were carried out on the positioning methods of direct coordinate regression and sub-region division.When the length of the sub-region is 2m,the average positioning accuracy can reach1.37 m,which is better than the direct positioning method of 2.66 m under the same network learning conditions.The indoor visual positioning method is integrated into the 3D indoor navigation system developed by the laboratory,and the results show that the usability and positioning accuracy of the system was significantly improved in the narrow and long linear public corridor scene common in large indoor buildings which greatly improves the user experience and has the value of promotion and application.