Enhancing Ball and Goal Detection with Image Preprocessing for Omni-Directional Vision

Real-time object detection using the YOLOv8 model on a soccer robot is often hampered by the unique visual challenges of omni-directional cameras, such as significant image distortion and high sensitivity to dynamic outdoor lighting variations. To address these issues, this study systematically investigates the effect of seven different image preprocessing techniques on improving ball and goal detection performance without requiring computationally expensive model retraining. Using the standard YOLOv8 model and an annotated outdoor dataset, different preprocessing categoriesincluding color correction, contrast enhancement, and blurring filters-are comparatively evaluated. Their performance is measured against a baseline using the mean Average Precision (mAP50) metric. The results show that White Balance, a color correction technique, significantly improves performance and achieves the highest mAP50 score (0.895) compared to the baseline (0.849). This success is attributed to its ability to neutralize the dominant environmental colors, resulting in more consistent intrinsic object features. In contrast, blurring-based methods prove counterproductive because they remove edge features crucial for detection. This study concludes that color correction is the most effective preprocessing strategy, offering a practical and efficient solution to improve detection robustness in dynamic environments.