Design of Weed Detection and Spraying Control System Based on Lightweight RT-DETR

Addressing the critical issues of low utilization rates and severe environmental pollution associated with traditional pesticide application methods, this paper proposes a realtime targeted spraying control system leveraging a lightweight Real-Time DEtection TRansformer (RT-DETR) model. In the weed detection module, high-precision real-time detection is achieved through two strategic improvements: a dynamic up-sampling-based feature pyramid and a streamlined encoder-decoder with operator fusion. For the spraying control module, a velocity estimation method based on Harris corner detection and pyramidal sparse optical flow is designed. Furthermore, a spraying timing model incorporating delay compensation is established, utilizing a multi-threaded parallel architecture to enable independent control of multiple nozzles. The system achieves integrated perception, decision-making, and control through deployment on the NVIDIA Jetson Orin Nano platform. Experimental results demonstrate that the improved model achieves a mAP@50 of 90.4% with an inference time of 35.2 milliseconds. Additionally, the end-to-end system latency is less than 100 milliseconds, fully satisfying the requirements for realtime spraying control.