Dynamic Recognition and Cutter Positioning Based on Morphological Features of Cane Tip Growth

Aiming to address the accuracy problem of cane tip recognition in complex natural environments, this paper proposes a cane tip feature annotation method based on the growth characteristics of sugarcane. In the context of the demand for lightweight and fast detection of cane tips, this paper optimizes the Yolov8n-Seg model with lightweight shared convolutional separated batch normalized detection head, model pruning, and knowledge distillation strategies. With these improvements, the accuracy of the optimized model increased by 0.2 percentage points, the number of parameters was reduced by 75.03%, the model size was reduced by 70.15%, the inference time is accelerated by 17.34%, and the GFLOPs were reduced by 40.00%. The lightweight cane tip detection model was deployed on the Jetson Orin NX platform with an average recognition frame rate of 7.42 f/s provides a lightweight hardware deployment solution for real-world applications in sugarcane harvesters. Finally, the depth camera was used for cane tip recognition and height measurement. The experimental results showed that the average relative errors of the camera were 0.189%, 0.675%, and 0.949% when the camera was 50 cm, 75 cm, and 100 cm away from the cane tip, respectively, which were all controlled within 1%, and were able to achieve accurate height measurement. Based on the statistical analysis of sugarcane clusters, this paper further proposes a sugarcane cluster identification method, providing a theoretical basis for saving adjustment time of the tip cutter during the harvesting process. It lays a theoretical and technical foundation for researching feature recognition, cutter height positioning, and real-time control of sugarcane harvester cuttings.