Research On Indoor Localization Algorithm Using Binocular Camera And Two-stream Convolutional Neural Network

With the developments of smart devices and people's living standards,the requirement for indoor localization services is getting increasingly stronger.Visual positioning techniques based on image information have developed rapidly in recent years.In order to take advantages of the independence and complementarity of gray image and depth image for binocular camera based indoor localization,in this paper,we study the indoor localization algorithm using the two-stream convolutional neural network(CNN)and binocular camera.The main contributions are as follows:(1)The theoretical knowledge of image based localization and CNN are studied.Some image based localization techniques and techniques are introduced at first.Then,the preliminary knowledge of CNN is given.Specially,the two-stream CNN algorithm is described in detail.At last,the network structure of two-stream CNN is built in the Tensor Flow platform.(2)A two-stream CNN based indoor localization algorithm using single binocular camera is proposed.In the off-line phase,after the image copy and image rendering process are used for the gray image and depth image,respectively,the three-channel grayscale and depth image is formed.Then,through the image cropping process,the obtained three-channel gray image and three-channel depth image are used for regression learning by two-stream CNN.In order to ensure the fusion effect,the weight parameters are calculated with the estimation error of each single mode image.At last,the position based regression function is obtained.In the on-line phase,after the image copy with the gray image and image rendering with the depth image,the final position can be estimated by the cropped image and regression function.The experiment results illustrated the efficiency of the proposed algorithm.(3)A two-stream CNN based indoor localization algorithm using multiple binocular camera is proposed.In the off-line phase,the gray image set and depth image set obtained from multiple binocular cameras are divided into three different image subset randomly.Then,the image stitching technique is used to form all gray images in each gray image set into a new gray image.Thus,three-channel gray image is obtained.Similarly,three-channel depth image is given.Next,through the image scaling,the two-stream CNN is used for regression learning and the position based regression function is obtained.In the on-line phase,through the corresponding image set division,image stitching and image scaling process of the obtained multiple gray images and depth images,the final position can be estimated with the regression function.The proposed algorithm makes full use of the gray and depth image information of multiple cameras to improve the localization performance.The experiment results illustrated that the proposed algorithm can obtain better localization performance than other existing algorithms.