Monitoring Of The Spraying Behavior In The Dragon Fruit Orchard Based On Human Posture Estimation And Scenes Interaction

The frequency of spraying behaviors in the greenhouse to a certain extent reflects the planting process and management methods of greenhouse crops.Therefore,achieving accurate recognition and classification of spraying behaviors is of great significance for intelligent monitoring and management of crop cultivation in the greenhouse.In response to the current problems of low real-time monitoring,poor transparency,and lack of systematic management of spraying,in this study spraying behavior recognition based on posture estimation and scene interaction in the dragon fruit orchard was proposed,with a view to providing a certain technical reference for the unmanned and intelligent development of monitoring in the planting process of installment agriculture.The main contents and relevant conclusions were as follows:(1)In view of the problem that the crops were distributed in a narrow way and the target detection method was not ideal in the dragon fruit monitoring,YOLACT model was used to segment crops and boxes.The results showed that the detection accuracy was 98%,the segmentation accuracy was 91.5%,and the average FPS was about 29.Moreover,the detection and segmentation effect in complex scenes was good,which indicated that the model was suitable for target detection and segmentation in facility environments,and has certain robustness.(2)Aiming at solving the problem of incomplete research on human behavior recognition in greenhouse in the agricultural field,a spraying behavior recognition model based on single person posture estimation and scene interaction was proposed to achieve recognition of human spraying behavior in video.Firstly,crop and boxes were detected based on YOLACT model,and the distance between them was calculated as the scene interaction feature.Then,the Open Pose model was used to recognize the posture and movement information of people as the human posture feature.Finally,the two were combined and support vector machine was used to achieve accurate detection of spraying behavior in greenhouse.The detection accuracy of this model was 90.57%,which could effectively distinguish spraying behavior from non spraying behavior;The recognition accuracy were 89.00%,87.63%,and 86.89% respectively under different lighting,block,and distance changes,with good robustness.The results showed that the model could be applied to accurately recognize spraying behavior in greenhouse,and could adapt to the requirements of spraying behavior recognition under different conditions.(3)Aiming at extracting the characteristics of videos in the dragon fruit orchard,such as fixed location,easy occurrence of personnel numbers,and complex and variable behavior,a model based on multi-person posture estimation and scene interaction was proposed to achieve recognition of multi-person spraying behavior in video.Firstly,the YOLOv5 s network was used to detect the boxes and humans,and then the human detection results were input into the HRNet network to detect its bone information.Secondly,the STGCN network was used to distinguish whether the bone information was a spraying action.Finally,combining the intersection ratio of the previous boxes and humans detection frame,it was determined whether the spray action occured.This model could effectively classify and identify spraying behaviors in experimental videos,with a detection ACC of 92.95%;The detection ACC of the algorithm were 92.33%,90.60%,92.74%,and 88.58% in various situations such as different personnel changes,different lighting,different degrees of block,and different distances,respectively.The algorithm could complete the recognition of multi-person spraying behavior in different situations,with good robustness.The experimental results showed that the model met the actual needs of complex and changeable scenes,and could be applied to recognize spraying behaviors in greenhouses.(4)Py Qt5 was applied to complete the GUI system design,combined with a spray behavior recognition algorithm based on multi-person posture estimation and scene interaction,a spraying behavior recognition system in greenhouse was designed.After the collection of video verification of the spraying behavior in the Pitaya greenhouse,the results showed that the system could accurately recognize the spraying behavior of multiple people,and met the functions of result visualization,recording,and backtracking.