A novel cascaded deep neural network for analyzing smart phone data for indoor localization

Abstract

This paper proposes and develops a cascaded deep neural network (CDNN) to analyze data, collected using the sensors of smart-phones, to accurately localize an object in an indoor environment. The indoor localization of an object (living or non-living) is important now a days, especially for those living alone or for securing valuable things. There are many existing studies that have attempted to identify the location of an inhabitant in a room through the analysis of the radio signal strength (RSS), with varying success. The strength of the RSS varies with distance and the presence of obstacles within the line of sight. As a result, an automated system using RSS signal in one environment might not work in another one. In this paper therefore, we propose and develop a different localization method based on data collected from different sensors embedded in a smart-phone. To analyze and predict the exact location within a very short distance (say a 1 to 1.5 m radius), we develop a novel CDNN. In this model, several deep neural networks (DNNs) are used in a tree structure where each node of the tree is an independent DNN. The output of any parent DNN at a specific node is used to decide which child node will be used to localize the position of an object. The combination of smart phone sensor data and the CDNN is applied in an academic building with an area of 175m2. The experimental results show that the proposed method can localize the subject/IoT object with a 74.17% accuracy within a 1.5 m radius and a 53% (approx.) accuracy within a 1 m radius. These performances are higher than those reported in many recent studies.