(An) efficient deep recurrent based indoor positioning system using mobile sensor data

Abstract

From the past few years, finding the accurate position of people, object or places has become thefocus of academia and industry. An accurate indoor positioning system gives rise to a number of usefulapplications. However, some challenges are associated to develop such system. First of all, data collectionis a major challenge. Recently, researchers have invented crowdsourcing and crowdsensing techniques tosolve the issues of manual data collection. These systems help to efficiently collect a large amount ofdataset but also raise the problem of noise in the data. Mobile sensors’ data is usually abstract in naturewhich means it is difficult to extract representative features from the dataset. Deep learning algorithmshave the potential to learn from such abstract data. The aim of this study is to find the optimal deeplearning approaches that can represent the mobile sensors’ data.

This study proposes a technique based on recurrent neural networks. The proposed system outperforms the other state of the art with respect to accuracy and reliability by finding the best hyperparameteres. Two-stage experiments were conducted for this evaluation, first on the Wi-Fi dataset forindoor positioning and second on the accelerometer and gyroscope dataset for detecting the short termstationary time periods. This zero velocity detection is done while walking and also while climbing stairs.The long short term memory network (LSTM) shows average error of only 2.23m. On the other hand,the zero velocity detection also shows a very good prediction accuracy. The reliability of the system istested by using an external dataset separately.

Through this study, the accuracy of the deep learning approaches in terms of mobile sensor data,derived parameters, model prediction, and technology acceptance is evaluated to check the potential ofdeep learning in real world indoor navigation. The results will be used to improve the system functionality, and to add the well-trained, suitable deep learning models in the indoor positioning field toovercome the discrepancies. Based on the findings, the usability of the system will be improved further.The practical implementation of the improved deep learning based indoor positioning system will notonly improve the quality of life but will also reduce the computational cost.