Improving Indoor Localization Using Convolutional Neural Networks on Computationally Restricted Devices

Indoor localization is one of the key enablers for various application and service areas that rely on precise locations of people, goods, and assets, ranging from home automation and assisted living to increased automation of production and logistic processes and wireless network optimization. Existing solutions provide various levels of precision, which also depends on the complexity of the indoor radio environment. In this paper, we propose two methods for reducing the localization error in indoor non line-of-sight (NLoS) conditions using raw channel impulse response (CIR) information obtained from ultra-wide band radios requiring no prior knowledge about the radio environment. The methods are based on NLoS channel classification and ranging error regression models, both using convolutional neural networks (CNNs) and implemented in the TensorFlow computational framework. We first show that NLoS channel classification using raw CIR data outperforms existing approaches that are based on derived input signal features. We further demonstrate that the predicted NLoS channel state and predicted ranging error information, used in combination with least squares (LS) and weighted LS location estimation algorithms, significantly improve indoor localization performance. We also evaluate the computational performance and suitability of the proposed CNN-based algorithms on various computing platforms with a wide range of different capabilities and show that in a distributed localization system, they can also be used on computationally restricted devices.