Leveraging Temporal Dynamics to Enhance Radar-Based Object Detection in Adverse Weather for Autonomous Vehicles

Autonomous vehicles require robust perception systems to detect and localize objects reliably under all environmental conditions. While cameras and LiDAR provide high-resolution perception, they suffer significant performance degradation in adverse weather. Radar sensors, with their inherent resilience to lighting variations and weather changes, present a strong alternative for robust perception. However, their effective use is challenged by higher noise, environmental clutter, and a wider beam-width that limits angular resolution. Recent work aims to address the limitations of radar-based object detection by exploring strategies such as raw radar data preprocessing, including denoising and resolution enhancement to improve the signal-to-noise ratio, and multi-modal sensor fusion to compensate for radar’s shortcomings. Radar-specific deep learning architectures have also emerged to enable more effective radar-only detection pipelines. While existing solutions have largely focused on spatial resolution enhancement, the temporal dimension remains significantly underexplored. Temporal modeling can provide critical benefits such as trajectory continuity, dynamic object discrimination, and resilience against transient noise and false reflections. Yet, most methods process radar data frame by frame, discarding valuable inter-frame dependencies due to challenges such as limited temporally annotated datasets and multi-frame alignment complexity. This dissertation addresses these gaps through a two-stage approach. First, a comprehensive study is conducted on the impact of different backbone networks and detection heads on radar-only object detection performance under normal weather conditions using the RADDet dataset. Second, a radar-only object detection framework is proposed, enhanced through self-supervised learning (SSL) to leverage temporal dynamics. Three SSL methods are adapted for 3D radar tensors: contrastive predictive coding (CPC), masked autoencoder (MAE), and time series to vector (TS2Vec). Evaluations on the K-Radar dataset demonstrate improvements over a baseline radar-only detector. MAE achieved the highest 3D sedan accuracy of 55.34%, while CPC achieved the best performance for buses and trucks with 3D accuracy rising from 43.98% to 51.29%. TS2Vec achieved 47.99% 3D sedan accuracy with minimal computational overhead. Regarding complexity, MAE was the most demanding, CPC compact but costly to train, and TS2Vec the most efficient. These results confirm that temporal SSL enhances radar perception in adverse conditions, offering a weather-resilient alternative to sensor fusion.

Name thesis:degree_name

Doctor of Philosophy (Ph.D.)

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Information Technology

Grantor dc:publisher

Carleton University

Year dc:date.issued

2025

Author dc:creator

• Sharifisoraki, Ziaaddin

Statement dc:rights

• Copyright © 2025 the author(s). Theses may be used for non-commercial research, educational, or related academic purposes only. Such uses include personal study, distribution to students, research and scholarship. Theses may only be shared by linking to the Carleton University Institutional Repository and no part may be copied without proper attribution to the author; no part may be used for commercial purposes directly or indirectly via a for-profit platform; no adaptation or derivative works are permitted without consent from the copyright owner.

Language dc:language.iso

en

DOI

10.22215/etd/2025-17005

OAI identifier oai:identifier

oai:carleton.scholaris.ca:20.500.14718/45162