Indoor Positioning System Using Regression-Based Fingerprint Method

Indoor positioning has become popular research topic in current days as this system can be used for diverse purposes. Compared with outdoor positioning like Global Positioning System (GPS), that did not work in indoor environment, indoor positioning using Radio Frequency (RF) gives possibilities in indoor environment with limitation of bandwidth needed depending how large the indoor environment (Terán, Aranda, Carrillo, Mendez, & Parra, 2017). Current challenge when using Radio Frequency is estimating position using signal strength received, because radio frequency has weakness like disturbances from human body that affect radio signals (Topak, Pekeriçli, & Tanyer, 2018). Others are uncertainty of signal strength due to fast frequency received that needs waiting time to show different position of moving person (Contreras, Castro, & de la Torre, 2017) and random behavior of received signal strength (Terán, Aranda, Carrillo, Mendez, & Parra, 2017).

Many technologies have been used. Starting with optical type (infrared (Santo, Maekawa, & Matsushita, 2017) and visible light communication (Zhang, Chowdhury, & Kavehrad, 2014)), sounds (Moutinho, Araújo, & Freitas, 2016) (Yayan & Yucel, 2015), and radio-frequency (Wi-Fi (Thuong, Phong, Do, Van Hieu, & Loc, 2016) and Bluetooth (Faragher & Harle, 2015) (Faragher & Harle, 2014). Among them, Bluetooth low energy (BLE) has been used frequently by reasons of low cost, very low battery consumption, and high availability as supported by most modern smartphones. However, each of them has different accuracy produced by estimating position, the accuracy depends on different aspects. Such as data preprocessing, localization method, or usage of machine learning or not. Many algorithms have been used for optimizing accuracy of the system. Such as multilateration and fingerprinting. Even so, there is not yet optimized solutions for high accuracy using BLE technology (Brena, et al., 2017).

There are many factor that affect the BLE radio propagation of the signals in indoor environments as BLE using radio signals, e.g., multipath effect, causing a random behavior in the Received Signal Strength (RSS) measurements caused by reflection (Terán, Aranda, Carrillo, Mendez, & Parra, 2017), movement rate of user (Topak, Pekeriçli, & Tanyer, 2018), and fast fading when measuring within a little time (Contreras, Castro, & de la Torre, 2017). To solve these problems, fingerprinting method is needed to estimate indoor position that needs estimation algorithm to ensure accuracy of position.

To get object’s location based on received signal strength from BLE, certain measurement method is needed. Current popular method is fingerprinting. Where localization algorithms used for measure or estimate location. It consists at least 2 phases: Offline phase and Online phase. Offline phase used to create a radio mapping of possible location from given signal strength received. While online phase (Brena, et al., 2017) will match the received signals during online moments with radio mapping from previous phase to determine object’s location. Older method to get object’s location using geometrical methods like trilateration (Rida, Liu, Jadi, Algawhari, & Askourih, 2015) to measure distance and position of person.

Many researchers tried to find optimized algorithm for indoor positioning. Some used k-nearest neighbor (Yu, et al., 2014) to estimate nearest points that can represent person’s position using classification, while certain researchers A few other tried regressions to estimate position like polynomial regression (Zhuang, Yang, Li, Qi, & El-Sheimy, 2016), where cumulative distribution function is related with error rate of average distance estimation. From these algorithms, regression model gives higher accuracy if compared with others. Given Polynomial Regression Model can solve multipath problems that gives random behavior of received signal strength (Zhuang, Yang, Li, Qi, & El-Sheimy, 2016).

Current state-of-the-art (Zhuang, Yang, Li, Qi, & El-Sheimy, 2016) is using Polynomial regression to calculate distance as propagation model. Where RSS received processed using weighted centroid localization or weighted sum to get coordinate and using polynomial regression model to get distance. Both are calculated using RSS signal from 3 advertisement channels of each beacon. then both results filtered using outlier detection to clean the result, by combine fingerprinting with polynomial regression model distance into combined distance. This filtered result will be processed using extended Kalman filtering using filtered distance from first outlier detection. Result from extended Kalman filtering will be filtered again using outlier detection to remove false measurement. This result then processed again with extended Kalman filtering into estimated position that will be compared with radio map to get the real position. This method is using distance-based measurement. Where the error rate is pretty high, caused by multipath effect and person movement rate, which is why the distance is filtered through many processes mentioned in the method. Other weakness is it takes lots of calculation time.

With previous facts mentioned before, most of the research did not use machine learning for indoor positioning system. Especially deep learning using convolutional neural network (CNN). Machine learning depends on the learning parameters and what kind of model used on the existing data, which poses new challenge of what kind of optimum parameters or model that produced high accuracy for locating position in indoor positioning system.

This research implements probabilistic method of fingerprinting using Deep Learning Convolutional Neural Networks Regression Model to estimate position of a person. The proposed method has shown to improve accuracy of estimated position. The design consists of BLE beacons as signal transmitter, and mobile smartphone as signal receiver. Signal received in the device is processed using fingerprinting method, then estimated by Convolutional Neural Networks (CNN) with self-designed architecture, resulting an estimated position of a person.

Researchers: Reginald Putra Ghozali dan I Gede Putra Kusuma Negara