Real-time recognition and dynamic positioning method for cotton terminal buds based on CottonBud-YOLOv5s algorithm and RGBD camera

摘要

• Developed CottonBud-YOLOv5s model for high-accuracy cotton bud recognition. • A visual positioning and dynamic compensation method for cotton terminal buds, named DAHEC, is introduced. • Achieved 97.9 % accuracy and 28.9 fps with CottonBud-YOLOv5s under challenging conditions. • Cotton topping machine maintained high recognition and localization accuracy at 0.2–0.8 m/s speeds. To address challenges such as motion blur, small target occlusion, and low real-time positioning accuracy during cotton mechanical topping operations, this study introduced the CottonBud-YOLOv5s lightweight cotton bud recognition model and Disjoint Area Visual Calibration (DAHEC) localization method. The recognition model incorporated the ShuffleNetv2 backbone network and DySample dynamic upsampling module to enhance real-time recognition performance. The model’s head and neck employed the ASFFHead detection head and global context (GC) mechanisms to improve the recognition accuracy under conditions of motion blur and small target occlusion. The DAHEC localization method combined camera imaging principles with depth information to convert the 2D cotton bud center coordinates into 3D camera coordinates, subsequently transforming them into robotic arm base coordinates based on the camera’s relative position to the robotic arm base. The method also dynamically compensated for cotton bud displacement caused by the mobile platform’s travel speed and model processing time. The experimental results indicated that the CottonBud-YOLOv5s model achieved a computational load of 4.40 G, an average accuracy of 97.9 %, a recall rate of 97.2 %, and a CPU recognition speed of 28.9 frames per second, outperforming other models in scenarios involving single plants, multiple plants, motion blur, and small target occlusion. Localization tests confirmed that the DAHEC method provided greater accuracy and stability than the TSAI localization method. The performance tests of the cotton topping machine indicated that at the driving speeds of 0.2–0.8 m/s, the topping rate ranged from 91.8 % to 96.9 %, the miss rate from 2.2 % to 4.6 %, and the false strike rate from 0.9 % to 3.6 %.