Research On Dynamic Recognition And Localization Method Of Cotton Top Bud Based On Unmanned Vehicle Platform In Field

Cotton is a vital strategic resource in China,playing a critical role in supporting domestic industrial development and ensuring the people’s standard of living.Topping cotton plants at an early stage of the natural growth cycle results in earlier boll formation and increased cotton yield,making it an essential aspect of the cotton industry.Currently,precise positioning technology is required for topping operations to achieve the desired operational outcomes.This study focuses on the accurate recognition and positioning technology of cotton top bud in the topping operation environment,proposes a dynamic recognition and positioning method for cotton top bud during the dynamic driving process on a field vehicle platform,and builds a complete hardware system based on an unmanned ground vehicle platform,aiming to promote the informationization and precision development of the cotton topping industry.The main research work and results are as follows:(1)Construction of a cotton top bud image dataset and selection of recognition algorithm.This study establishes a comprehensive cotton top buds image dataset in the MS COCO 2017(Microsoft Common Objects in Context 2017)format,which includes seven cotton plants’top buds under four lighting conditions and two shooting distances over three time periods.In this study,Faster R-CNN,Cascade R-CNN,Retina Net,and Auto Assign were used for cotton top bud detection experiments.The experimental results show that Faster R-CNN has an average precision 0.3%,2.6%,and 3.1%higher than Cascade R-CNN,Auto Assign,and Retina Net,respectively,demonstrating a superior detection accuracy and better adaptability to cotton top bud field detection scenarios.(2)Research on a cotton top bud recognition method in the field based on an improved Faster R-CNN.This study proposes an improved Faster R-CNN cotton top bud detection model to address the detection difficulties brought about by complex field backgrounds,lighting conditions,and the small volume of cotton top bud.Reg Net X-6.4GF,Guided Anchoring(GA),Generic Ro I Extractor(GRo IE),and Dynamic R-CNN mechanisms were introduced to optimize the feature extraction network,anchor generation mechanism,region of interest extraction mechanism,and positive and negative sample allocation mechanism of Faster R-CNN.The experimental results show that the improved Faster R-CNN achieves an average precision of 98.1%in cotton top bud detection,demonstrating strong robustness and good generalization ability.High-precision detection of cotton top buds has been realized.(3)Dynamic recognition and positioning method of cotton top bud in the field based on an unmanned vehicle platform and system construction.This study proposes a cotton top shoot identification and positioning method suitable for dynamic driving processes of field-carried platforms and builds a complete hardware system based on an unmanned vehicle platform.The hardware system consists of an Intel Real Sense D435i depth camera,an unmanned vehicle platform,a DELTA three-axis parallel manipulator,and a notebook industrial control computer.The method is based on the unmanned vehicle platform completing dynamic identification and positioning of cotton top buds through RGB-D information,deep learning,spatial coordinate transformation,and dynamic coordinate correction technologies.In this study,Intel Real Sense D435i was used to capture images of cotton top buds,and a dataset of cotton top shoots required by the system was constructed.To meet the real-time requirement for dynamic scene recognition,this study selects the Cascade R-CNN as the recognition algorithm with slightly lower accuracy than Faster R-CNN but with speed advantages,and uses Res Ne St-50 as its feature extraction network.Meanwhile,Side-Aware Boundary Localization(SABL)mechanism is integrated to improve recognition accuracy.(4)Verification experiments of the cotton top bud dynamic recognition and positioning system based on an unmanned vehicle platform.The validation experiments were divided into detection accuracy experiments for an improved Cascade R-CNN algorithm,static positioning accuracy experiments for the system,and dynamic positioning accuracy experiments.The experimental results show that the improved Cascade R-CNN algorithm achieves an average detection accuracy of 99.3%.In the static experiment,the average absolute errors of the system in the the X_w-axis,Y_w-axis,and Z_w-axis of the world coordinate system were within 5.3mm,3.1mm and 3.3mm,respectively.In the dynamic experiment,under the speed state of less than 0.3m/s,the average absolute errors of the system in the the X_w-axis,Y_w-axis,and Z_w-axis of the world coordinate system were 5.3mm,6.8mm and 3.4mm,respectively.The experimental results demonstrate that the proposed method in this study can achieve high-precision dynamic recognition and positioning of cotton top bud.In summary,the proposed method and hardware system for dynamic recognition and positioning of cotton top bud based on an unmanned vehicle platform in this study can achieve high accuracy recognition and positioning,and have the feasibility for application.They can contribute to the promotion of the informatization,intelligence,and precision development of the cotton industry.