基于RGB-D 感知与点云碰撞检测的禽蛋采收方法

Poultry Egg Harvesting Method Based on RGB-D Perception and Point Cloud Collision Detection

摘要:

散养模式下，禽蛋采收机器人存在识别精度不足和抓取过程碰撞中断等问题。本文以YOLO v8n-OBB模型为基础，引入MobileNet轻量化结构并提出方向感知增强的SimAM注意力机制，同时，结合基于分布焦点损失的旋转角度精细回归改进，进一步提升了禽蛋锚框定位精度与角度回归准确性。相较于YOLO v8n-OBB模型，改进模型精确率从97%增长到98.5%，平均精度均值从98.9%增长到99.1%，对禽蛋角度推理准确度从86.3%增长到91.2%，模型参数量缩减13%，能够很好地适配边缘计算设备的硬件资源限制。在对禽蛋识别与定位后，通过深度相机生成环境点云，针对机械臂姿态的3个旋转自由度，采用差异化策略生成角度候选序列，对候选姿态下的工具点云以及环境点云进行碰撞检测，得到最终的有效无碰撞吸取姿态。经点云处理后，系统点云碰撞检测机制稳定生效，而单枚禽蛋碰撞检测平均耗时缩短至0.23 s。在实际吸取场景中，禽蛋总体吸取成功率达78%。本文方法适用于禽蛋采收机器人禽蛋快速识别与定位，有效解决了视觉感知系统在性能受限设备上部署，以及实际场景中可能存在碰撞的技术挑战，构建了禽蛋采收全过程的高性能鲁棒模型，为智慧型自动化农业生产流程的进一步发展提供理论支持与实践参考。

Abstract:

In free-range farming, poultry egg harvesting robots face problems, including insufficient recognition accuracy and task interruption due to collisions during grasping. An improved model was developed based on the original YOLO v8n-OBB framework. A MobileNet-style lightweight architecture was introduced, and a direction-aware enhanced SimAM attention mechanism was proposed. Meanwhile, a refined rotation angle regression strategy based on distribution focal loss was adopted to further boost the localization accuracy of poultry egg bounding boxes and orientation regression precision. Compared with the original YOLO v8n-OBB model, the improved model increased the precision from 97% to 98.5% and the mAP from 98.9% to 99.1%. Meanwhile, the orientation inference accuracy of poultry eggs rose from 86.3% to 91.2%, and the model parameter volume was reduced by 13%. It can well adapt to the hardware resource constraints of edge computing devices. Following poultry egg recognition and localization, environmental point clouds were generated via a depth camera. A differentiated strategy was then employed to generate candidate angle sequences for the three rotational degrees of freedom of the robotic manipulator. Collision detection was conducted between the tool point clouds under candidate poses and the environmental point cloud to determine the final effective collision-free suction pose. After point cloud processing, the system maintained stable point cloud collision detection accuracy, and the average collision detection time for a single poultry egg was shortened to 0.23 s. In practical harvesting scenarios, the system achieved a poultry egg recognition success rate of 99% and an overall harvesting success rate of 78%. The proposed method was suitable for rapid poultry egg recognition and localization for egg harvesting robots, effectively solving the technical challenges of deploying visual perception systems on performance-limited devices and avoiding collisions in real-world scenarios. It constructed a high-performance and robust model for the whole poultry egg harvesting process, providing theoretical support and practical reference for the further development of intelligent automated agricultural production systems.