Design and implementation of a lightweight vision system for edge-computing AGVs

Abstract

To overcome the trade-off between constrained on-board computation of edge devices like AGVs in smart warehousing and logistics and the high real-time demands of visual perception, this paper proposes a lightweight object-detection system based on the Jetson Nano. The ultra-lightweight YOLOv5n network is chosen as the detector; a special dataset is gathered to train the model on unstructured industrial-logistics scenes; and Mosaic data augmentation is applied during training to enhance performance in challenging environments. To reduce the impact of the compute and memory-bandwidth limitations of the target platform, we propose a co-optimization approach that uses lower-resolution input, TensorRT operator fusion, and FP16 half-precision quantization. Our experiments demonstrate that the proposed system is significantly more robust to lighting variations and multiple targets than a traditional OpenCV pipeline. On the Jetson Nano, the fine-tuned model achieves mAP@0.5 = 99.5 % on the test set, with single-frame inference-only latency decreased from 44.42 ms to 23.17 ms and inference-only throughput of 43.83 FPS. The model has 1.9 M parameters, which provides a good trade-off between compute, power, and real-time performance on the edge device, and a consistent deployment pattern for the visual subsystems of industrial mobile robots.