Deep Learning-Based Human Position Estimation in Complex Environments Using Millimeter Wave Radar Signals

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu | Master's thesis
Date
2023-06-12
Department
Major/Subject
Control, Robotics and Autonomous Systems
Mcode
ELEC3205
Degree programme
AEE - Master’s Programme in Automation and Electrical Engineering (TS2013)
Language
en
Pages
65+1
Series
Abstract
Currently, mobile robots are quite common in factories and warehouses. However, most of them are not able to detect targets around corners as their sensors mainly work in line-of-sight environments. It poses a potential risk to the workers, as they may have critical collisions with robots around corners. An alarm system may help, but it is not user-friendly for individuals with hearing disabilities. Thus, a robot would better prevent collisions if it could locate and track people around corners. Significant work has been done to develop feasible methods for localizing people around corners. However, most of them cannot work in complex environments or require well-established infrastructure. Thus, data-driven methods show great potential to solve this problem because of their ability to handle complex and high-dimensional data. In this thesis, the aim is to present a novel and promising solution for the detection and localization of targets located around corners. The proposed approach has been validated through real-world experiments. The core of the solution involves a radar system consisting of multiple embedded system boards that capture radio frequency data, complemented by a ground truth system utilizing cameras to provide accurate localization labels. Before feeding the data into the localization convolutional neural network, we employ advanced preprocessing techniques such as the multiple signal classification algorithm and linear interpolation to refine the raw data. To enhance the accuracy of our system in accurately positioning individuals, we leverage either the Kalman filter or the particle filter. These filters are applied to the preliminary results obtained from the convolutional neural network. By incorporating these filtering techniques, we achieve exceptional accuracy and robustness in determining the exact positions of individuals around corners.
Description
Supervisor
Särkkä , Simo
Thesis advisor
Tan, Bo
Keywords
convolutional neural network, frequency-modulated continuous-wave radar, none-of-sight tracking, multiple signal classification
Other note
Citation