Research On Strawberry Detection And Positioning Method Based On Improved YOLOv5x

With the rapid development of Smart and Digital Agriculture,the demand for efficient,fast and accurate agricultural production methods is also increasing.Manual strawberry picking is the most time-consuming and labor-intensive part of the production process,and the workload is not only large,but also the operating period is long,so the use of strawberry picking robots to achieve automatic picking is a research direction with practical value.However,some strawberry robot prototypes have problems such as low detection rate and inaccurate positioning for strawberry fruits in complex environment and different posture in traditional greenhouse.Therefore,it is necessary to research an effective method to improve the accuracy and stability of strawberry fruit detection and positioning.This thesis mainly focuses on the strawberries from the planting base in Changfeng County,Anhui Province,and proposes an improved method for detecting strawberry fruits using SD-YOLOv5 x.In addition,the study introduces the use of a depth camera to locate the picking points of strawberry fruits.The main research content is as follows:(1)Data collection and pre-processing for the strawberry dataset.The dataset of strawberries was collected in the field at the strawberry growing base in Changfeng County.8300 images were extracted and methods including cropping,geometric transformation and data enhancement were used to expand the images to over 12,000,enhancing the size and diversity of this strawberry dataset.The strawberry images were annotated using annotation software to classify the strawberries into green,white ripening and red ripe states according to their growth stage,and converted into a dataset format suitable for the YOLO network.(2)Strawberry fruit detection based on SD-YOLOv5 x model.To address the problem that the traditional target detection algorithm is not accurate in strawberry detection under the complex environment of greenhouses,this article improves the YOLOv5 x model by embedding CCH feature extraction module and NAM attention mechanism mainly in Backbone network and Neck network,respectively.In the comparison experiments with the Faster R-CNN,YOLOv4 and YOLOX algorithm models,their m AP is improved by 2.56%,8.78% and 6.58%,respectively.Compared with the original YOLOv5 x network model,it has higher detection accuracy and higher classification confidence for strawberry fruits with3%,0.81% and 2.54% improvement in m AP,precision and recall,respectively.(3)Research on 3D strawberry fruit localization method based on Azure Kinect DK depth camera.To address the problem of inaccurate localisation of strawberry fruits by traditional target detection algorithms,the depth information measurement principle of the depth camera and the conversion relationship of four coordinate systems are introduced,and repeated localisation and effective distance verification experiments are conducted to test the accuracy and stability of the depth camera.Using Zhang’s camera calibration technique,the depth camera was calibrated and error evaluated.Finally,the improved detection model was used to detect strawberry fruits,and the 3D coordinates of the picking point of ripe strawberry fruits were obtained by finding the central position of the detection frame,thus completing the spatial localization process of strawberry fruits.