基于OWD-YOLO的火龙果成熟度识别与三维姿态估计方法

Dragon Fruit Maturity Classification and 3D Pose Estimation Method Based on OWD-YOLO

摘要:

针对复杂果园环境下火龙果成熟度识别与姿态估计困难导致采摘精度不高的问题，以YOLO 11n-pose为基础，提出一种轻量高效的OWD-YOLO实时检测模型，实现火龙果精准、高效采摘。在基础模型中引入重参数化卷积，并与C3K2模块结合，增强对火龙果多尺度特征和细节姿态特征的提取能力;在SPPF模块中引入小波池化与大核卷积注意力机制，减少光照变化、背景遮挡等环境因素干扰，提高模型检测精度;在主干网络部分引入DGECA注意力机制提升模型对火龙果果皮颜色和纹理等重要特征的识别能力，改善成熟度判别准确性;在复杂果园环境中部署基于OWD-YOLO模型的六自由度机械臂采摘机器人平台，并通过深度相机实现火龙果三维姿态估计。田间试验结果表明，OWD-YOLO模型目标检测精确率、平均精度均值与关键点检测平均精度均值分别达到88.0%、92.7%、93.3%，较基础模型分别提高5.0、2.4、2.0个百分点，平均帧率为58.7 f/s，单果采摘成功率达86.0%，平均采摘时间为29.4 s，能够满足复杂果园环境下精准机械化采摘需求。

Abstract:

In response to the challenges of low harvesting accuracy due to difficulties in the maturity recognition and pose estimation of dragon fruits in complex orchard environments, a lightweight and efficient OWD-YOLO real-time detection model was proposed based on YOLO 11n-pose to achieve precise and efficient harvesting of dragon fruits. Firstly, reparameterized convolution was introduced into the base model, combined with the C3K2 module, to enhance the model's ability to extract multi-scale features and fine-grained pose details of the dragon fruit. Secondly, by incorporating wavelet pooling and large kernel convolution attention mechanisms into the SPPF module, the model reduced interference from environmental factors such as lighting variations and background occlusion, thereby improving detection accuracy. Additionally, a DGECA attention mechanism was introduced into the backbone network to enhance the model's ability to recognize key features such as the fruit skin color and texture, improving the accuracy of maturity classification. Finally, a six-degree-of-freedom robotic arm harvesting platform based on the OWD-YOLO model was deployed in a complex orchard environment, with three-dimensional pose estimation of the dragon fruit achieved via a depth camera. Field experiments demonstrated that the OWD-YOLO achieved an object detection precision of 88.0%, a mean average precision of 92.7%, and a keypoint mean average precision of 93.3%, with absolute improvements of 5.0, 2.4, and 2.0 percentage points over the baseline, respectively. The average frame rate was 58.7 f/s, with a single fruit harvesting success rate of 86.0%, and an average harvesting time of 29.4 s. These results met the requirements for precise mechanized harvesting in complex orchard environments.