DePos: accurate orientation-free indoor positioning with deep convolutional neural networks

The smartphone-based indoor positioning has attracted considerable attention in recent years. In order to implement accurate and infrastructure-free positioning systems, researchers have tried to fuse magnetic field, Wi-Fi, and dead reckoning information applying particle filter technique. In fact, magnetic signals have high-resolution and Wi-Fi signals are able to provide coarse-grained global results. However, in order to move particles, the particle filter requires the phone's orientation aligned with the user moving directions, thus limiting its applications and impairing user experiences. In order to implement an orientation free and infrastructure-free system, we propose a deep learning based positioning scheme. The proposed system constructs a new kind of rich-information positioning image, then leverages convolution neural network to automatically map positioning images to position predictions. We also present a novel extracting and labeling method to generate enough positioning images for training the neural network. Finally, experiments convincingly reveal that the proposed positioning system is orientation-free, infrastructure-free, and achieves good precisions.