Authors

Abstract

In this study, a machine-vision-based automatic control system for the topsoil stripping of Fritillaria ussuriensis Maxim. (FUM) was designed to address the problems of manual adjustment, low control accuracy, and response lag in stripping-depth control during FUM harvesting. An improved YOLOv5s-SA target detection algorithm was used to calculate FUM density and was deployed on the Jetson Nano edge-computing platform. Combined with a fuzzy control algorithm, it drives the servo electric cylinder to achieve dynamic depth adjustment of the scraping board. Test results showed that, after deploying the target detection algorithm on the edge AI device and accelerating it with TensorRT, the average inference time was 0.077 s, and the system response time was 0.26 s, meeting the real-time requirements of agricultural operations. Simulation results indicated that the average error between the stripping depth of the automatic control system and the preset depth was 3.72 mm, representing a 44.1% improvement compared with fixed-depth control. The average ideal stripping rate reached 54.96%, an improvement of 21.66% over the 33.3% achieved under fixed-depth control.

Abstract in Chinese

针对平贝母收获中表土剥离深度依赖人工调节、控制精度低及响应滞后等问题，设计了一种基于机器视觉的平贝母表土剥离自动控制系统。采用改进的YOLOv5s-SA目标检测算法实现平贝母识别密度计算，并部署于 Jetson Nano 边缘计算平台上，结合模糊控制算法驱动伺服电缸，实现刮土板的动态深度调节。所提出的控制系统试验结果表明，目标检测算法在边缘AI设备部署并经TensorRT加速后，识别平均推理耗时为0.077 s，响应时间0.26s，满足农业作业实时性要求。模拟试验结果表明，自动控制系统剥离深度与预铺设深度平均误差为3.72 mm，相较固定深度控制改善44.1%，平均剥离理想率为54.96%。相比于固定深度控制理想剥离率33.3%，提升了21.66%。