Enhancing Road Object Detection in Fisheye Cameras: An Effective Framework Integrating SAHI and Hybrid Inference

Abstract

Fisheye cameras are extensively employed in surveillance systems because they provide a broad viewing angle, enhancing visibility. The reception of an image from a wide perspective can result in distortion, posing challenges for recognition systems, mainly when dealing with moving objects, as observed in traffic systems. This work presents an effective framework comprising multiple modules to address the issue of small objects and rapidly changing viewing perspectives in fisheye camera data. First, we use Slicing Aided Hyper Inference (SAHI), an algorithm that uses generic slicing-aided inference to deal with small objects. Second, we integrate the outcomes of CNN (YOLO) and state-of-the-art Transformer (Co-DERT) detection methods to utilize the respective strengths of each strategy for handling data limitations. This approach has demonstrated promising performance, achieving an F1 score of 0.6077 and achieving the 4<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> in Track 4 of the AI City Challenge 2024.