Corn stem diameter and ear orientation angle measurement method based on D3-YOLOv11 and RGB-D camera

The operational effect of the reverse ear picking device for fresh corn is affected by stem diameter and ear orientation angle. The existing devices lack the ability to sense these parameters in real-time, making it difficult to dynamically adjust operating parameters, which leads to a high damage rate and harvest loss. To this end, this study focuses on the visual perception aspect and proposes a recognition method based on a depth camera and an improved D3-YOLOv11 segmentation model, which provides reliable visual input for subsequent adaptive regulation. Specifically, this study proposes Dual-Domain Dynamic Gate Conv (D3GConv) to enhance the multi-scale feature extraction ability of the model. In the neck network, a bidirectional weighted pyramid structure with semantic detail injection is designed to improve the segmentation accuracy of small objects. Generalized Focal Loss V2 was used to optimize the detection head to enhance the accuracy of boundary localization in dense stem scenes. Finally, the depth information is fused to realize the real-time measurement of stem diameter and ear orientation angle. Experimental results show that the Mask-mAP50 of the D3-YOLOv11 model reaches 99.3% and 94.6% in stem and ear instance segmentation tasks, respectively. The Mean Absolute Error of stem diameter measurement based on depth information is only 0.16 cm, and the Coefficient of Determination of ear orientation angle reaches 0.95, which verifies the reliability and practicability of this method in the adaptive control of the ear harvesting device. It provides an effective visual perception basis for improving the intelligence level of equipment.