DeepNavi

Abstract

Indoor navigation plays a crucial role in indoor location-based services. Single signal-based navigation systems, however, are prone to sensor noises, signal ambiguities and are specific to trial sites. To address these, existing work fuses different signals with user trajectories. Despite their accuracy, many of them are specific to input signals and navigation modes (e.g., spot-based or sequence-based) and are computationally expensive in large sites. Additionally, they do not give predictive uncertainty estimations, leading to a lack of trust in navigation instructions. In this paper, we propose a unified framework for accurate indoor navigation in various modes with different inputs, termed DeepNavi. We exploit either convolutional or recurrent neural networks for initial feature extraction. Afterwards, we insert fully connected layers to generalize extracted signal-dependent features to a shared domain before fusion. Then, we leverage state-of-the-art ensemble learning to learn multiple predictive models. By combining them together, we further reduce the impact of signal noises and achieve high accuracy. Finally, we insert mixture density networks to model more generalized data distributions and provide uncertainty estimations. We have implemented DeepNavi and conducted extensive experiments in two different trial sites with different signal combinations. Experimental results show that DeepNavi reduces location errors by more than 20% with comparable orientation accuracy.