INDOOR POSITIONING USING CONVOLUTION NEURAL NETWORK TO REGRESS CAMERA POSE

Abstract. In recent years, the issue of indoor positioning has become more and more popular and attracted more attention. Under the absence of GNSS, how to more accurately position is one of the challenges on the positioning technology. Camera positioning can be calculated by image and objects. Therefore, this study focuses on locating the user's camera position, but how to calculate the camera position efficiently is a very challenging problem. With the rapid development of neural network in image recognition, computer can not only process images quickly, but also achieve good results. Convolution Neural Network (CNN) can sense the local area of the image and find some high-resolution local features. These basic features are likely to form the basis of human vision and become an effective means to improve the recognition rate. We use a 23-layer convolutional neural network architecture and set different sizes of input images to train the end-to-end task of location recognition to regress the camera's position and direction. We choose the sites where are the underground parking lot for the experiment. Compared with other indoor environments such as chess, office and kitchen, the condition of this place is very severe. Therefore, how to design algorithms to train and exclude dynamic objects using neural networks is very exploratory. The experimental results show that our proposed solution can effectively reduce the error of indoor positioning.