Multi-Source Fusion Localization Technology Based on Convolutional Neural Networks

Abstract

The fingerprint localization technology based on receive signal strength (RSS) is one of the most important methods in indoor localization. However, the stability of RSS is fragile due to internal multipath effects. This problem causes the reference points (RPs) to be mis-matched during system localization. Subsequently, the localization performance is degraded. To reduce the fluctuation of signal, we use a WiFi/ Bluetooth Low Energy (BLE)/ Pedestrian Dead Reckoning (PDR) fusion localization system, but the problem that RPs are mis-matched still exists. Therefore, in this paper, we propose a WiFi/BLE/PDR fusion localization system based on convolutional neural network (CNN) to solve the problem. The method makes use of the signal features that can not be captured by the traditional positioning systems, and improves the localization accuracy. In the training phase, the scheme utilizes RSS to construct RP fingerprint sequences and set up a CNN structure. WiFi and BLE localization models are trained with the structure. In the test phase, the trained models are employed for localization. Then, the location information from WiFi, BLE, and PDR is combined to obtain the location coordinates. The result shows that the average error of the method is 2.47 m. Compared with traditional methods, the method proposed in this paper is more effective.