Accurate and Stable Wi-Fi based Indoor Localization and Classification Using Convolutional Neural Network

Abstract

Listen

Translate article

Indoor localization is essential for providing location based services inside homes, malls, and hospitals. Wi-Fi routers are available in almost every building and Wi-Fi chipsets are also available in almost every smartphone. Therefore, fingerprinting of Received Signal Strength Indicator (RSSI) values coming from Wi-Fi routers is a cheaper option for indoor localization. In conventional Wi-Fi fingerprinting methods, RSSI values are collected at various indoor locations and stored in a database. The device which needs localization, collects new RSSI values from its current unknown location. These values are compared with the database and the best match is returned as the current user location. Due to differences in Wi-Fi chipsets and environmental conditions, RSSI values fluctuate which makes accurate, stable, fast, and precise determination of user location difficult. If the user is inside a large multi-floor building, dataset scalability and RSSI fluctuations can make the task even more difficult. User tracking and determination of the direction in which the user is moving also becomes challenging due to hurdles and non-walkable points in the indoor environment. To solve these issues, in this paper we present a Wi-Fi fingerprinting method for large indoor environments that uses 1-D convolutional neural networks (CNN) for floor and region-status (hurdle, walkable point) classification. The procedure consists of collecting RSSI dataset which is then normalized and pre-processed. This step is essential for training the classification and localization model. The trained model can be used in real-time for fast, stable, and accurate classification of floors, region-status and user location coordinates. Based on our experiments inside a two floor university library, the proposed approach can classify the floors and region-status with an accuracy of 70.50% and 81.23% respectively, while the mean localization error is 3.47 m.