Answers are grounded in this paper and cite their sources.
Visual perception systems are essential for harvesting robots in smart agriculture, but deployment is often limited by computational constraints. For real-time truss tomato detection in complex greenhouses, existing models rarely deliver high accuracy, low latency, and lightweight design on resource-constrained edge devices, especially under variable illumination. We introduce PHDT-DETR, a lightweight, end-to-end detector optimized for edge deployment. Building on the RT-DETR baseline, PHDT-DETR integrates a CSP-PMSFA backbone for efficient multi-scale feature extraction, a CA-HSFPN neck that enhances feature fusion via Coordinate Attention, a DRBC3 block that enhances multi-scale feature representation through multi-branch re-parameterized convolutions while trimming redundant computation, a TS-IFI encoder that reduces attention complexity, and a joint NWD+Shape-IoU regression loss that provides overlap-independent, aspect-ratio–aware supervision for slender, irregular tomato skewers. We further apply Layer-Adaptive Magnitude-based Pruning (LAMP) for aggressive compression. Experiments show that the pruned model achieves 90.8% mAP50 while reducing the parameter count to 6.1 M and the computational cost to 17.4 GFLOPs. Deployed on an NVIDIA Jetson Orin Nano Super and compiled with TensorRT, the model runs at 66.0 FPS with a compact 15.5 MB footprint, outperforming mainstream YOLO models. These results demonstrate the feasibility of deploying high-precision, real-time, end-to-end object detectors on resource-constrained edge devices for robotic harvesting in greenhouses.
No takes yet. Share an insight, caveat, or question.
yang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: