Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Back

9 Views

ICTD 2025

Data Sensing & Analytics

Advanced Sensing Applications in Traffic Operations, Safety, and Data Analytics

Optimized Generative Adversarial Network for roadside LiDAR-based vehicle detection

Monday, June 9, 2025

3:15 PM – 3:30 PM MT

Primary Author(s)

Shanglian Zhou, PhD

Postdoctoral Research Fellow
University of Hawaii at Manoa
Honolulu, Hawaii, United States

Presenter(s)

Shanglian Zhou, PhD

Postdoctoral Research Fellow
University of Hawaii at Manoa
Honolulu, Hawaii, United States

Co-Author(s)

Igor Lashkov, PhD

Postdoctoral Research Fellow
University of Hawaiʻi at Mānoa
Honolulu, HI, United States
Hanyi Yang, Ph.D

Postdoctoral Research Assistant
University of Hawaii at Manoa
Honolulu, Hawaii, United States
AN

Amanda Nitta, n/a

Graduate student
University of Hawaiʻi at Mānoa
Honolulu, Hawaii, United States
Guohui Zhang, PhD

Professor
University of Hawaiʻi at Mānoa
Honolulu, Hawaii, United States

Submitter(s)

Shanglian Zhou, PhD

Postdoctoral Research Fellow
University of Hawaii at Manoa
Honolulu, Hawaii, United States

Abstract: With deep learning-based technologies rapidly emerging as a prominent approach for transportation research, one prerequisite is the vast availability of annotated data for model training. For vehicle detection using roadside Light Detection and Ranging (LiDAR) data, one bottleneck is the lack of publicly available large-scale datasets that contain complex traffic scenarios at road intersections. Roadside LiDAR data refer to the data collected using LiDAR sensors installed on transportation infrastructure at road intersections, such as traffic light poles. Although there exist various autonomous driving LiDAR datasets, the different characteristics between autonomous driving and roadside LiDAR data make it difficult to directly apply deep learning models to learn the domain knowledge from autonomous driving scenarios for detecting traffic objects under roadside scenarios. This paper proposes a deep learning framework for roadside LiDAR-based vehicle detection, by leveraging a Generative Adversarial Network (GAN) to generate synthetic vehicle point clouds for model training. To promote the performance of GAN, an automated optimization approach through Bayesian optimization is proposed to jointly determine the training hyperparameters as well as the architecture layout of GAN. With the optimal GAN obtained, the generated vehicle point clouds are used to augment roadside LiDAR data to construct a large training dataset for deep learning-based vehicle detection. Through a comparative study, a series of YOLO models are implemented and trained on the synthetic LiDAR dataset, to validate the efficacy of the proposed method in boosting deep learning model performance on roadside LiDAR-based vehicle detection through synthetic LiDAR data.

Learning Objectives:

Attendees can expect to learn the following from this session:

Upon completion, participant will be able to understand the mechanism of Generative Adversarial Networks (GANs) in generating synthetic LiDAR data.
Upon completion, participant will be able to learn the proposed optimization strategy based on Bayesian optimization to search for the optimal training hyperparameters as well as network layout.
Upon completion, participant will be able to understand the effect of GAN in enhancing deep learning model performance on roadside LiDAR-based vehicle detection through synthetic LiDAR point clouds.