Multi-Target spraying behavior detection based on an improved YOLOv8n and ST-GCN model with Interactive of video scenes

Intelligent monitoring and management of pesticide spraying behavior significantly affects the growth of crops and product quality. In this study, a multi-target spraying behavior detection with an interaction of video scene in Pitaya orchard was proposed. Firstly, the improved YOLOv8-n model was utilized for human and sprayer target detection. Secondly, the top-down AlphaPose keypoint detection model was utilized to detect human skeletal keypoints. Then, the ST-GCN model was employed to explore the action features on time axis, and then, the model that could recognize spraying actions was obtained. Finally, the sprayer target and human spraying behavior were fused to achieve accurate detection of multi-target spraying behaviors. To verify the effectiveness of the algorithm, a collected 60-minute video was tested. The results showed that the mAP@50 of the improved YOLOv8-n model was 99.5%. The detection accuracy and speed of the AlphaPose key detection model were 73.3% and 30.98 fps, respectively. For simultaneous detection of multi-person spraying, the Accuracy of Classifier (ACC) was 95.48%. Furthermore, an effectiveness analysis was conducted on different situations, such as: occlusion, distance variation, and lighting in multi-person spraying, the average accuracy of 89.71% was obtained. All the results showed that the method could be utilized to identify multi-person spraying behavior in facility environments, which could provide technical reference for standardization of fruit traceability system and agricultural product quality supervision.