Enhancing object detection for humanoid robot soccer: comparative analysis of three models

Abstract

The humanoid robot soccer system encounters a notable challenge in object detection, primarily concentrating on identifying the ball and often neglecting crucial elements like opposing robots and goals, resulting in on-field collisions and imprecise ball shooting. This study comparatively evaluates three you only look once (YOLO) real-time object detection system variants: YOLOv8, YOLOv7, and YOLO-NAS. A dataset of 2104 annotated images, covering classes such as ball, goalpost, and robot, was curated from Roboflow and robot-captured images. The dataset was partitioned into training, validation, and testing sets, and each YOLO model underwent extensive fine-tuning over 100 epochs on this custom dataset, leveraging the pre-trained common objects in context (COCO) model. Evaluation metrics, including mean average precision (mAP) and inference speed, assessed performance. YOLOv8 achieved the highest accuracy with a mAP of 0.92, while YOLOv7 showed the fastest inference speed of 24 ms on the Jetson Nano platform. Balancing accuracy and speed, YOLO-NAS emerged as the optimal choice. Thus, YOLO-NAS is recommended for object detection for humanoid soccer robots, regardless of team affiliation. Future research should focus on enhancing object detection through advanced training techniques, model architectures, and sensor fusion for improved performance in dynamic environments, potentially optimizing through scenario-specific fine-tuning.