Carleton University
From ODD Definition to Deployment: Weather-Resilient Perception in Autonomous Driving via Sensor Benchmarking, Adaptation, and Fusion
Abstract
dc:description.abstractWhile modern autonomous driving perception systems achieve strong performance under normal environments, their reliability can deteriorate in adverse weather conditions. In the autonomous vehicle (AV) workflow, degraded object detection may lead to safety-critical failures, such as failing to detect pedestrians or cyclists in low visibility, mislocalizing vehicles at longer ranges, or producing spurious detections from backscatter and sensor noise that destabilize downstream planning. This thesis investigates how object detection should be designed and evaluated to remain reliable within operating conditions. To make reliability explicit and testable, we adopt an operational design domain (ODD) framework and treat the ODD as an interface between environmental conditions and perception requirements. In this framework, we define key ODD parameters relevant to adverse-weather detection to structure evaluation and report performance per-ODD slice rather than as a single metric. Within this ODD framework, first, we conduct a cross-modality study of perception sensor performance using AV modalities, including RGB cameras, thermal cameras, and LiDAR. We quantify adverse-weather degradation across conditions and object scales, analyze failure modes, and show that no single modality is robust across illumination regimes, motivating ODD-aware sensor selection and fallback design. In addition, given deployment constraints, including latency and cost, we develop an efficient single-sensor approach, FAA-YOLO, a camera-based detector that adapts its inference to reduced visibility caused by fog. FAA-YOLO serves both as a deployable baseline and as a fallback under restricted ODD operation when sensing conditions deteriorate. However, sensor-level evidence reveals coverage gaps that motivate multi-sensor fusion for improved robustness. To enable reliable fusion, we introduce a camera-LiDAR calibration pipeline tailored to low-resolution LiDAR-camera configurations, improving geometric consistency in challenging scenarios. Building on this calibration foundation, we propose a bird’s-eye view (BEV) fusion method augmented with label-guided knowledge distillation, which helps the runtime-aware student model learn robustness from a high-capacity teacher model suitable for real-time operation. Finally, we consolidate models, calibration methods, and empirical evidence developed across the ODD definition, sensor evaluation, single-sensor adaptation, calibration, and fusion into ODD-aware deployment guidance. This includes recommendations for sensor selection, fallback, restricted-ODD policies, monitoring strategies, and acceptance criteria for weather-resilient perception in autonomous driving.
Degree
thesis:*- 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
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Abbasi, Hasan
Subjects
dc:subject × 3Rights
dc:rights- Statement dc:rights
-
- Copyright © 2026 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