E-YOLO to OMOD: An Enhanced YOLO Framework for Small Moving Object Detection in Omnidirectional Videos

Omnidirectional cameras provide a comprehensive 360-degree field of view and enable efficient data collection, making them valuable for various applications such as surveillance, autonomous navigation, and traffic monitoring. However, challenges such as image distortion, low resolution, and the difficulty of detecting small objects remain significant obstacles. YOLO, a widely adopted object detection model, offers high-speed and accurate recognition but struggles with small object detection in omnidirectional videos due to its grid-based approach, which can result in feature loss in low-resolution images. To address these issues, this study improves the detection accuracy of moving objects in omnidirectional images by leveraging transfer learning. A new dataset was carefully curated from omnidirectional camera footage, with particular emphasis on annotating key feature regions of each moving object to reduce the impact of resolution constraints. By incorporating deep learning-based enhancements, the proposed approach significantly improves the recognition and classification of small moving objects, particularly in outdoor environments such as intersections. Experimental results demonstrate that the proposed method outperforms conventional methods, offering greater robustness and reliability for real-world applications. The proposed method achieved an accuracy of 90%, whereas conventional models achieved a maximum accuracy of only 53%.