CRNet: Corner Recognition from Trajectories Based on Convolutional and Recurrent Neural Networks

Corner recognition plays an important role in trajectory-route matching for fingerprint crowdsourcing in indoor positioning systems. However, the pose diversity and fake corner problem in most practical scenarios could much degrade the recognition performance. This paper deals with the two problems from a deep learning perspective. We design the CRNet consisting of convolutional and recurrent neural networks, which can recognize a corner from pedestrian movement trajectories directly from raw sensor measurements. Field experiments have validated its effectiveness in terms of much improved precision, recall and F1-measure compared with the state-of-the-art schemes. As the CRNet does not require manual feature extraction, we envision its wide applications in practical corner recognition scenarios.

I. Introduction

In the last decade, fingerprinting has been extensively researched for indoor localization systems, where the basic idea is based on the assumption that each indoor location can be represented by a unique feature called fingerprint [1]–[4] In fingerprinting-based indoor localization systems, crowdsourcing approaches have been recently adopted to relieve or even eliminate the burdensome site survey for radio map construction by exploiting causally collected fingerprints [5]–[8]. Though not collected at specified locations, crowdsourced fingerprints still need to be annotated with some location information for radio map construction [9]–[13]. To this end, some researchers have proposed to extract crowdsourced fingerprints from a pedestrian movement trajectory. As long as a trajectory can be correctly matched to one physical route, each step position can be obtained from the floorplan to annotate the corresponding step fingerprint in the trajectory [14]–[17].