Normal view MARC view ISBD view

Lidar object perception and classification using 3d machine learning algorithm / Ericsson Yong

By:

Ericsson, Yong [author.]

Material type: Text

text

Media type:

unmediated

Carrier type:

volume

ISBN:

THE0010425 (Local)

Subject(s):

Dissertation note: Thesis (Master of Science) -- Universiti Malaysia Pahang - 2025 Abstract: Autonomous vehicles (AVs) are emerging technologies aimed at reducing the rate of ever-increasing traffic accidents at Malaysia involving heavy vehicles. One of the key technologies in detecting obstacles during autonomous vehicle navigation is 3D LiDAR sensor which captures sparse and disordered point clouds making identifying objects inside it challenging. To address this, deep learning has been utilized in point clouds and now it can be represented in several forms like pillar, voxel or point form for feature extraction in object detection. In this research, 3D LiDAR data of heavy vehicles are collected using both a stationary setup (static) and a mobile platform (dynamic) with a sensor mounted on top of a moving vehicle within a university campus. Then, 3D object detectors were pretrained using transfer learning with the addition of ‘Van’ and ‘Truck’ class. Also, their anchor dimensions are based on the most frequently used van and truck models in Malaysia. The 3D object detectors selected are PointPillars, PartA2Free and SECOND to represent the pillar-form and the voxel form respectively. Upon pretraining them, there’s are changes in mAP (mean Average Precision) scores for the 3D object detectors, especially for the Van and Truck class. The same pretrained 3D object detectors are then deployed onto point cloud scenes and are compared between the original detectors in terms of classification score. It was observed that the pretrained 3D object detectors had an improvement on recall and a reduction in False Negative detections, confirming that transfer learning enhances detection of objects. These findings prove that the dataset collected for transfer learning is feasible to improve object or vehicle recognition in Malaysian Autonomous Vehicles applications.

Tags from this library: No tags from this library for this title. Log in to add tags.

Average rating: 0.0 (0 votes)

Holdings ( 2 )
Title notes ( 4 )

Holdings
Item type	Current library	Call number	Status	Date due	Barcode
Thesis	UMPLIB PEKAN	FTKPM .Y66 2025 r Thesis (Browse shelf(Opens below))	Not for loan		T000004800
Thesis	UMPLIB PEKAN	CD14001 (Browse shelf(Opens below))	Final Processing		T000004801

Faculty of Manufacturing and Mechatronic Engineering Technology

Thesis (Master of Science) -- Universiti Malaysia Pahang - 2025

Includes bibliographical references

Autonomous vehicles (AVs) are emerging technologies aimed at reducing the rate of ever-increasing traffic accidents at Malaysia involving heavy vehicles. One of the key technologies in detecting obstacles during autonomous vehicle navigation is 3D LiDAR sensor which captures sparse and disordered point clouds making identifying objects inside it challenging. To address this, deep learning has been utilized in point clouds and now it can be represented in several forms like pillar, voxel or point form for feature extraction in object detection. In this research, 3D LiDAR data of heavy vehicles are collected using both a stationary setup (static) and a mobile platform (dynamic) with a sensor mounted on top of a moving vehicle within a university campus. Then, 3D object detectors were pretrained using transfer learning with the addition of ‘Van’ and ‘Truck’ class. Also, their anchor dimensions are based on the most frequently used van and truck models in Malaysia. The 3D object detectors selected are PointPillars, PartA2Free and SECOND to represent the pillar-form and the voxel form respectively. Upon pretraining them, there’s are changes in mAP (mean Average Precision) scores for the 3D object detectors, especially for the Van and Truck class. The same pretrained 3D object detectors are then deployed onto point cloud scenes and are compared between the original detectors in terms of classification score. It was observed that the pretrained 3D object detectors had an improvement on recall and a reduction in False Negative detections, confirming that transfer learning enhances detection of objects. These findings prove that the dataset collected for transfer learning is feasible to improve object or vehicle recognition in Malaysian Autonomous Vehicles applications.