A Location-Tracking Method With a Convolutional Neural Network

Introduction

In recent years, GPS (Global Positioning System) has been mainly used to get location information of various devices such as automobiles or smart phones. GPS can identify the location of a device by receiving multiple satellite signals with its receiver (Bassma, G., 2018). In general, it has 2 to 5 meters of error length (Michael, G. W., 2005). However, we would not be able to get location data precisely indoors, such as a factory or a shopping mall, since the connection to the GPS would be affected by obstacles. There are several ways of getting location information indoors. Using acceleration sensors or beacon devices are well-known methods. However, the former has a disadvantage due to accumulating errors. The latter can communicate within a radius from 30m to 100m. It cannot figure out the precise location even with the range. Once the device lost connection to the beacon server, it cannot do anything for the decision to move.

When using a robot running in a factory or a shopping mall, it is important to confirm the current location of the robot. The robot can run a predefined course to help the robot to move to a destination by a line tracer. In contrast with this, it is challenging to run a given course in a shopping mall because it needs to avoid obstacles such as people or carts.

Then, the authors postulate a scene that auxiliary such as a line trace cannot be used. This article will find a way to identify the location of the robot until it reaches the destination. As a method, camera images are analyzed by using a deep learning technique and then developing a system to identify the location and orientation indoors. The condition to acquire the location information indoors is to use the camera image without using the sensor and not connect with other devices. Specifically, a camera on the robot takes many pictures while moving. After that, the robot learns the pictures by the deep learning engine and identifies the location and orientation of the robot.

This paper is an extended version of a paper submitted to the 9th International Workshop on Web Services and Social Media (WSSM-2020), held in conjunction with the 23rd International Conference on Network-Based Information Systems (NBiS-2020). The improvements of this paper from the paper (Kawakami, S., 2020) consists as follows:

• •

  Considered the mark to improve the robustness against the difference of the angle to the mark.

• •

  Found the learning method that we can achieve 97% accuracy by trying some methods.

This paper is constructed into five sections. In the introduction section, we introduce our motivation and proposed location tracking method briefly. In the background section, we show the background of this research field. In the experiment description section, we describe our experimental environment, applied convolutional neural network (CNN), and six types of discriminators to clarify the information type suitable for location tracking using only an image. In the experimental results section, we will discuss the experimental results. Finally, we conclude this paper in the conclusion section.

Background

Radio Frequency Identifier (RFID) has been introduced as a related technology for acquiring the location information of people and objects indoors (Youngbo, L., 2011). Yann L. et al. proposed a system that uses RFID to track people indoors to protect assets in confidential areas. Using RFID location tags, the accuracy of the system to inform people entering the confidential area was 84% (Aiguo, L., 2016). RFID can identify targets and collect certain digital information by using radio waves. Compared with other locating technologies, RFID locating system has many advantages such as low cost and low energy. However, there are some disadvantages, such as security issues and network configuration issues.