YOLOv10-pose and YOLOv9-pose: Real-time strawberry stalk pose detection models

Abstract

In the computer-aided industry, particularly within the domain of agricultural automation, fruit pose detection is critical for optimizing efficiency across various applications such as robotic harvesting, aerial crop surveillance, precision pruning, and automated sorting. These technologies enhance productivity and precision, addressing challenges posed by an aging labor force and the increasing demand for sophisticated robotic applications in agriculture. This is particularly crucial for strawberries, which are globally recognized for their high nutritional value. The strawberry pickting robots generally cut the stems, so knowing the pose of the strawberry stalks before cutting can effectively adjust the pose of the end effector, thereby improving the success rate of picking. This paper referred to the keypoint detection branch and loss function of the YOLOv8-pose model, and combined the latest YOLOv9 and YOLOv10 object detection models to propose YOLOv9-pose and YOLOv10-pose. The experimental results showed that YOLOv9-base-pose had the best comprehensive performance, reaching 0.962 in Box_mAP50 and 0.914 in Pose_mAP50, and the speed met the real-time requirement of FPS 51. The entire YOLOv10-pose series did not achieve satisfactory accuracy, but not using non-maximum suppression did indeed speed up the post-processing. In the YOLOv10-pose series, YOLOv10m-pose achieved the best comprehensive performance with Box_mAP50 of 0.954, Pose_ mAP50 of 0.903, and a speed of 61 FPS. Comparing YOLOv9-base-pose with the entire series of YOLOv8-pose and YOLOv5-pose also demonstrated the superior performance of YOLOv9-base-pose. YOLOv9-pose and YOLOv10-pose can provide a theoretical basis for pose detection and a reference for other similar fruit pose detection.