University of Leeds
Camera-LiDAR Fusion-based 3D Object Detection for Autonomous Driving
Abstract
dc:description.abstractThis thesis addresses the challenge of robust 3D object detection for autonomous driving through advanced integration of camera and LiDAR data. Reliable environmental perception is fundamental to autonomous systems, yet remains hindered by occlusion, truncation, sensor sparsity, and dynamic real-world conditions. Although multi-modal fusion methods have shown promise, existing approaches often exhibit limited cross-modal interaction and rigid fusion strategies, resulting in suboptimal performance in complex scenarios. To overcome these limitations, this research proposes a flexible, adaptive camera–LiDAR fusion framework designed to enhance both feature representation and fusion flexibility. First, a hierarchical multi-modal feature fusion network is developed to enable deep interaction between low-level geometric cues and high-level semantic features, improving the exploitation of complementary information across modalities. This design significantly enhances detection performance, particularly for distant and partially occluded objects. Building on this, a context-aware dynamic fusion mechanism is introduced to adaptively re-weight sensor contributions based on their reliability under varying conditions. This enables the model to respond robustly to environmental uncertainty and reduces false detections caused by degraded sensor inputs. The proposed framework is extensively evaluated on the KITTI benchmark, achieving consistent and notable improvements over state-of-the-art baselines in 3D Average Precision across all difficulty levels. Experimental results demonstrate that the combination of hierarchical feature interaction and adaptive fusion substantially improves robustness in challenging scenarios, including long-range detection and severe occlusion. Overall, this thesis makes two key contributions: (1) a novel hierarchical cross-modal fusion architecture that strengthens feature representation, and (2) a dynamic, context-aware fusion strategy that enhances adaptability to real-world conditions. These contributions provide both theoretical and practical advancements in sensor fusion, supporting the development of more reliable and scalable perception systems for autonomous driving.
Degree
thesis:*- Name dc:type.qualificationname
- M.Phil
- Level dc:type.qualificationlevel
- masters
- Grantor dc:publisher.institution
- University of Leeds
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hou, Yiming
- Advisors dc:contributor.advisor
-
- Markkula, Gustav
- Li, Wenjing
Rights
- Language dc:language.dcterms_iso639_2
- eng