Greenhouse Ventilation Equipment Monitoring for Edge Computing

Abstract

Digital twins based on real-world scenarios are heavily reliant on extensive on-site data, representing a significant investment in information technology. This study aims to maximize the capabilities of visual sensors, like cameras in controlled-environment agriculture, by acquiring more target-specific information at minimal additional cost. This approach not only reduces investment but also increases the utilization rate of existing equipment. Utilizing YOLOv7, this paper introduces a system with rotatable pan-tilt cameras for the comprehensive monitoring of large-scale greenhouse ventilation systems. To mitigate the computational load on edge servers at greenhouse sites caused by an abundance of video-processing tasks, a Region of Interest (ROI) extraction method based on tracking is adopted. This method avoids unnecessary calculations in non-essential areas. Additionally, we integrate a self-encoding approach into the training phase, combining object detection and embedding to eliminate redundant feature extraction processes. Experimental results indicate that ROI extraction significantly reduces the overall inference time by more than 50%, and by employing LSTM to classify the state of the fan embedding sequences, a 100% accuracy rate was achieved.

Keywords

• YOLOv7
• autoencoders
• state recognition
• edge computing