Object-based Indoor Localization using Region-based Convolutional Neural Networks

Abstract

This study presents a novel method that utilizes Region-based Convolution Neural Network (R-CNN) for identifying objects in an image, and then, localizes the image using a pre-stored map of these objects. The algorithm follows the next steps. First, an image, taken by a smart phone, is uploaded to the server, where a pre-trained R-CNN and template images are combined to identify the objects in the image. In the second step, the location of the image is derived according to the spatial relationships of the objects in the image with a prior knowledge of the objects location in a digital map. Finally, if more than one location solutions are obtained from the previous step, a range-based localization method is utilized to eliminate the wrong results. To prove the concept, a test is conducted in typical yet challenging indoor environment. Selected objects were labeled in 600 images, taken randomly around the test environment, to train the R-CNN and used for template matching. Another 42 images, with known capture locations, were utilized to assess the performance of the method. The proposed object-based indoor localization method detected 81.7% of the objects in the images, and provided 59.5% success rate with 1-5 m accuracy.