Real-Time UAV Patrol Technology in Orchard Based on the Swin-T YOLOX Lightweight Model

Abstract

Using unmanned aerial vehicle (UAV) real-time remote sensing to monitor diseased plants or abnormal areas of orchards from a low altitude perspective can greatly improve the efficiency and response speed of the patrol in smart orchards. The purpose of this paper is to realize the intelligence of the UAV terminal and make the UAV patrol orchard in real-time. The existing lightweight object detection algorithms are usually difficult to consider both detection accuracy and processing speed. In this study, a new lightweight model named Swin-T YOLOX, which consists of the advanced detection network YOLOX and the strong backbone Swin Transformer, was proposed. Model layer pruning technology was adopted to prune the multi-layer stacked structure of the Swin Transformer. A variety of data enhancement strategies were conducted to expand the dataset in the model training stage. The lightweight Swin-T YOLOX model was deployed to the embedded platform Jetson Xavier NX to evaluate its detection capability and real-time performance of the UAV patrol mission in the orchard. The research results show that, with the help of TensorRT optimization, the proposed lightweight Swin-T YOLOX network achieved 94.0% accuracy and achieved a detection speed of 40 fps on the embedded platform (Jetson Xavier NX) for patrol orchard missions. Compared to the original YOLOX network, the model accuracy has increased by 1.9%. Compared to the original Swin-T YOLOX, the size of the proposed lightweight Swin-T YOLOX has been reduced to two-thirds, while the model accuracy has slightly increased by 0.7%. At the same time, the detection speed of the model has reached 40 fps, which can be applied to the real-time UAV patrol in the orchard.

Keywords

• model lightweight
• real-time UAV patrol orchard
• layer pruning
• deployment model