A Novel Approach to Object Detection Under Diverse Illumination Conditions Using Row-Wise Exposure Images

In this study, we present a novel approach to object detection utilizing row-wise exposure (RWE) images to substantially improve object detection performance in low- and high-illumination conditions. Unlike previous RWE imaging techniques that require row-wise merging for high dynamic range (HDR) synthesis, our system directly utilizes raw RWE images without postprocessing. It enables the proposed object detection algorithm to effectively overcome the limitations of conventional systems in dynamic and variable lighting scenarios. Additionally, we introduce a tailored data augmentation strategy optimized for the unique characteristics of RWE images, enhancing model training and robustness without relying on dedicated RWE datasets. We developed a prototype CMOS image sensor (CIS) with RWE functionality to demonstrate the practical viability of our approach. This prototype was instrumental in validating the system's effectiveness in real-world conditions. The proposed data augmentation method, designed specifically for raw RWE images, enriches the training dataset, enabling models to adapt to various lighting situations and improve their generalization abilities. Our experiments employed widely adopted object detection models, such as YOLOv7 and YOLOv9, along with standard datasets, such as MS COCO and HDR4RTT, to evaluate model performance under varying illumination coefficients and motion blur intensities. The results showed significant performance improvements over existing object detection methods, especially in challenging illumination conditions and dynamic environments. To ensure reproducibility and allow further validation, we have made our source code available at: https://github.com/eomtaehoon/RWE-YOLO.