Indoor Wi-Fi Localization Based on CNN Feature Fusion Network

With the rise of 5G new smart city construction, the demand for location-based services (LBS) has been increasing rapidly. Indoor positioning technology based on Wi-Fi has attracted extensive attention due to its advantages of low deployment cost and high positioning accuracy. However, traditional neural networks ignore a large amount of available information in the intermediate layer when conduct feature extraction of Wi-Fi signal data, resulting in poor localization performance and robustness. In order to solve this drawback, this paper proposes a novel convolutional neural network (CNN) feature fusion network which considers both spatial features and intermediate layer features. Specifically, it normalizes the raw data by z-score to reduce the impact of data fluctuation. Then the spatial features are extracted using CNN and a flatten layer is added after its pooling layer to extract the intermediate layer features. Finally, all features are merged into the fully connected layer. The experimental results show that our proposed fusion network outperforms existing localization algorithms.

I. Introduction

Indoor positioning technology based on Wi-Fi has attracted extensive attention due to its advantages of low deployment cost and high positioning accuracy. It mainly include two categories: ranging-based positioning methods, e.g., Time Of Arrival (TOA), and non-ranging positioning methods, e.g., fingerprinting. Compared with range-based methods, fingerprint positioning methods are widely used because of their simple implementation[1]. However, fluctuations in received signal strength indicator (RSSI) can have a significant impact on localization accuracy, so fingerprint localization is often combined with machine learning or deep learning to obtain better robustness. Convolutional neural network (CNN), as a typical deep learning network, has been widely utilized. It usually uses a convolution-pooling structure and a general filter to extract features[2]. S. Aikawa et al. modeled the adjacency relationship between APs as a two-dimensional model and used it to construct a CNN model for fingerprint localization[3]. G. Cerar et al. constructed an improved CNN model based on channel state information (CSI) from MIMO[4]. However, although the middle layer has large amount of available information, all these existing researches ignore this advantage. Hence, we propose a fusion network considering both spatial features and intermediate layer features, which fully utilizes features to contribute the localization performance.