Human Localization Based on 3D LiDAR Point Cloud Utilizing Equifunctional Deep Learning Approaches

Budiyanta, Nova Eka (2024) Human Localization Based on 3D LiDAR Point Cloud Utilizing Equifunctional Deep Learning Approaches. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7022211004-Dissertation.pdf - Accepted Version Restricted to Repository staff only until 1 April 2027. Download (4MB) | Request a copy

Abstract

The tradition of ascertaining human position through image data from cameras significantly infringes upon an individual’s privacy. Moreover, in the domain of computer vision, the utilization of picture data from cameras
cannot yield precise distance measurements but is confined to estimation. In this context, an alternate data source that addresses privacy concerns and provides robust spatial data is the 3D LiDAR point cloud. The data is acquired from the reflection of numerous LiDAR lasers that operate rotationally and continuously with a 360◦ horizontal field of view. Nonetheless, obstacles exist in the processing of 3D LiDAR point cloud data. The constraints in data representation are attributed to the inherently scattered and unstructured character of point cloud data, particularly
that of 3D LiDAR. The representation of the item is limited to its surface, which serves as the plane of reflection for the LiDAR laser.

This study aims to present a solution to the issues associated with processing 3D LiDAR point cloud data for the
purpose of detecting human position represented in 3D LiDAR point cloud environment. The present investigation employs two primary methodologies for data analysis as quifunctional deep learning approaches: structured point processing and direct point processing. During the preliminary phase of the study, a feature extraction procedure was conducted to eliminate planar surfaces from the 3D LiDAR point cloud data
representation.

The outcomes of the feature extraction procedure were subsequently utilized to evaluate the efficacy of the human localization method. This was accomplished by contrasting the accuracy and speed outcomes of structured point processing with those of direct point processing approaches. Structured point processing yielded an optimal model with an accuracy of 71.66% and an inference time of 12.2 ms, whereas direct point processing achieved an accuracy of 98.79% with an inference speed of 37 ms. This study effectively acquired 3D LiDAR point cloud human data and subsequently implemented
distance determination to accurately localize human objects inside the raw 3D LiDAR point cloud data, facilitating future research on human activity monitoring.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Human Localization, 3D Point Cloud, LiDAR, Structured Point Processing, Direct Point Processing, Deep Learning
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5105.546 Computer algorithms T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Nova Eka Budiyanta
Date Deposited:	09 Jan 2025 07:02
Last Modified:	09 Jan 2025 07:02
URI:	http://repository.its.ac.id/id/eprint/116224

Actions (login required)

View Item