A method for identifying grape stems using keypoints

Abstract

In grape farm operations, reliable methods to locate stems among grapes are often needed. In this study, we propose a top-down stem localization method using grape growth characteristics. The method, which is based on human pose estimation, uses a target detection algorithm to identify grape clusters and applies a keypoint detection model to locate stems. We use different target detection algorithms and keypoint detection models in turn for validation, and combine the advantages of a Ghost module and HRNet to build a new lightweight HRNet. A comparison of different keypoint detection models demonstrates that YOLO v5 target detection algorithm achieves high detection accuracy for grape clusters, with an average accuracy of approximately 92%. The final experimental results reveal that grape bunch stems can be detected effectively. The best results are achieved with YOLO v5m as the detection algorithm using a lightweight HRNet, with a 90.2% stem recognition accuracy. Furthermore, MobileNetV3 is the fastest, with a detection speed of 7.7 frames per second. Finally, we discuss the advantages and disadvantages of our method compared with the traditional stem detection method, prospects for assistance with machine picking, production management, and so on.