Machine Learning based Framework for Target Tracking and Detection

  • Lana Hasan Alhaj Hussain

Student thesis: Master's Thesis

Abstract

This thesis addresses the issue of tracking a mobile target, through means of a data-driven active sensor selection mechanism in the Internet of Things (IoT) sensing applications. IoT networks have proved their usefulness in environmental monitoring applications, including target tracking, thanks to their ease of deployment, scalability, and ability to provide real-time monitoring. Nevertheless, IoT-based target tracking approaches usually sacrifice either the accuracy of the approach or the energy conservation, and are applied in applications where the target and its location(s) are generally known. In this work, we present a predictive target tracking approach, capable of efficiently tracking a mobile target that has an unknown location and intensity, with high accuracy and low overall energy cost. The proposed approach utilizes Deep Learning (DL) techniques to accurately predict the target's location by learning highly complex and non-linear relationships between the input and output. The predicted location is used to dynamically select a set of sensors that can carry out continuous tracking. Convolutional long short-term memory (CNN-LSTM) model is proposed to predict the mobility of a source using time-delayed data readings obtained from the IoT sensors. Moreover, a mechanism to detect possible target loss and its subsequent recovery is also proposed. An example of tracking a radioactive source carried by a walking person in varying settings is used to validate the proposed approach performance. An IoT radiation sensor readings dataset based on models of radiation physics, and derived from both a real-life human mobility dataset and a synthetically generated dataset, is utilized for training a predictive model. The proposed approach shows that it can accurately predict the target's location and carry out effective tracking while saving up to 90% of the network's energy when using a CNN-LSTM model to predict the target's location.
Date of AwardJul 2022
Original languageAmerican English

Keywords

  • Sensor Selection
  • Deep Learning
  • CNN-LSTM
  • Radiation Tracking
  • Target Tracking.

Cite this

'