Date of Award

6-7-2026

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2026

Department

Computer Science and Engineering

First Advisor

David Anastasiu

Abstract

Effective traffic management is hampered by the physical limitations of traditional pinhole cameras, whose narrow fields of view create blind spots that require costly multi-camera infrastructure to overcome. Fisheye cameras offer a wide-angle alternative capable of single-camera intersection coverage, but the radial distortion inherent in their optics renders standard object detection algorithms unreliable, limiting their practical deployment in traffic surveillance.

This thesis investigates specialized object detection pipelines for fisheye traffic imagery by reproducing and building upon two top-ranked submissions from the 2025 AI City Challenge Track 4: the VNPT and SKKU-AutoLab pipelines. Both were successfully reconstructed outside their original containerized environments on an HPC cluster, achieving F1 scores of 0.613 and 0.622 respectively. Two improvement methods were then developed and evaluated against the Fisheye8K benchmark. The first applied training modifications — including area-proportional loss weighting, an additional detection head, and radial augmentation strategies — combined with weighted box fusion post-processing, achieving a +1.8% mAP improvement over baseline on the external evaluation set. The second pursued a data-centric reconstruction using a newer model architecture and synthetic weather augmentation, which performed competitively on the local validation set but generalized poorly to unseen data, falling below the baseline on external evaluation.

These results demonstrate that post-processing ensembling provides the most reliable gains under dataset scarcity, while synthetic augmentation alone is insufficient to bridge the generalization gap without access to more diverse real-world fisheye data.

Download

Share

COinS