Detection-Driven Key-Frame Edge Offloading for Real-Time UAV Grape Bunch MOTS

Abstract

Real-time instance-level perception is required for UAV operations in vineyards, but on-board compute and rural uplinks preclude full-frame streaming and large models. We develop an edge-offloading pipeline in which a TensorRT-accelerated YOLOv8n on a Jetson Xavier NX performs detection-driven key-frame selection and region-of-interest (ROI) compression, while multiobject tracking and segmentation (MOTS) runs on an edge server. The onboard selector filters redundant frames using Intersection-over-Union matching. A three-thread parallelized schedule overlaps detection, selection, and compression. On two vineyard sequences (frontal and path-planned), FP16 TensorRT reduced detector latency by 48% (to 27.5 ms). The proposed offloading strategy discarded 54–69% of redundant frames, reducing bandwidth consumption to 23–27% via ROI-centric compression. While the end-to-end throughput ($\approx$ 1.3 FPS) is bounded by the detector, it supports online monitoring at typical vineyard flight speeds. Compared to a full-sequence baseline, our selective offloading significantly reduced identity switches (IDS) (from 40 to 6 in frontal views) and improved relative tracking accuracy by stabilizing association, improving tracking stability (notably reducing IDS) relative to a full-sequence baseline under bandwidth-constrained field operations, while absolute MOTS remains challenging in cluttered vineyard scenes.