Date of Award

6-9-2025

Document Type

Thesis - SCU Access Only

Publisher

Santa Clara : Santa Clara University, 2025

Department

Computer Science and Engineering

First Advisor

David Anastasiu

Abstract

Object detection plays a crucial role in traffic surveillance for road safety and traffic management. Road Object Detection can be used for monitoring traffic flow or traffic analysis. In traffic monitoring systems, fish-eye cameras are particularly useful because they can cover larger areas of streets and intersections, reducing the need for multiple cameras. Their ability to provide wide, omnidirectional coverage, which traditional cameras with limited fields of view (FoV) cannot offer, is the reason for fish-eye camera’s recent popularity.

However, these cameras also introduce image distortion, requiring complex techniques for undistortion and unwarping, or specialized processing methods to manage the distortion effectively. The AI City Challenge 2024, Track 4 introduces a novel fish-eye camera dataset for the 2D road object detection task, FishEye8K.

Many previous methods rely on ensembles with different combinations of YOLO and transformer models under Weighted Box Fusion (WBF). These techniques are also coupled with image enhancement and super-resolution models to handle images taken at night and low-resolution images respectively. In addition to the image distortion, there is a lack of open fish-eye image datasets for road object detection, with. To combat this, previous techniques generate data through augmenting the VisDrone dataset to be create synthetic fish-eye images. Despite these complex methods, the model performance of the previous top team, VNPT AI, results in an F1-score of 0.6406.

This paper presents a detection framework leveraging RTMDet with an ensemble of predictions to improve robustness against fisheye distortions. We train this new model on both the FishEye8K and augmented VisDrone datasets and incorporate the popular WBF ensemble method. Our combined approach achieves notable results, with an F1-score of 0.6413.

Download

Share

COinS