【摘要】 针对青贮饲料收获抛送环节人工操控抛送筒劳动强度大、现有算法在动态作业场景下落料点检测精度低、计算复杂度高等问题，本文提出了一种面向田间复杂工况的青贮饲料抛送落料点检测方法。首先，在YOLO v8n-pose模型基础上，通过引入轻量坐标注意力机制(LCA)、动态卷积(DynamicConv)及可变形卷积(DCNv4),提升料厢角点与青贮饲料落料点检测精度；其次，融合RGB-D数据，配准像素点与点云中的对应点，基于凸包算法拟合料厢边缘；最后，通过叉积法计算料厢与落料点的包含关系。试验结果表明，改进YOLO v8n-Pose模型在青贮饲料流摆动松散状态下mAP50:95达到95.1%,较原始模型提升5.9个百分点，在青贮饲料流正常状态下mAP50:95达到95.0%,较原始模型提升3.3个百分点。改进模型在不同青贮饲料流状态下均表现出更高的检测精度与稳定性，尤其对形态变化剧烈的异常工况适应能力显著增强，为自适应抛送控制提供了视觉检测基础。更多还原

【Abstract】 The current method for loading material onto vehicles in silage harvesters, which primarily involves manually controlling the rotation of the throwing arm, is labor-intensive and has high operational requirements. This method not only affects the harvesting efficiency but also easily causes losses of silage. A detection method for determining whether silage fell into the trailer hopper was proposed. Firstly, an improved YOLO v8n-pose model was constructed. By introducing the lightweight coordinate attention(LCA), DynamicConv, and Deformable Convolution v4(DCNv4), the detection accuracy of the trailer hopper corners and the falling points of silage was improved. After coordinate transformation, combined with the convex hull algorithm and cross product method, it was determined whether the silage fell into the trailer hopper. Experiments proved that the improved YOLO v8n-Pose model achieved a mAP50:95 of 95.1% in the loose state of the silage flow, an increase of 5.9 percentage points compared with that of the original model, and a mAP50:95 of 95.0% in the normal state of the silage flow, an increase of 3.3 percentage points compared with that of the original model. The improved model demonstrated higher detection accuracy and stability in different states of the silage flow, with a significant enhancement in adaptability to abnormal working conditions with drastic shape changes, which provided a solid visual foundation for future adaptive throwing control.更多还原