Row Segmentation and Local Path Planning in Fruit Plantations using Lightweight Neural Networks

Título: Row Segmentation and Local Path Planning in Fruit Plantations using Lightweight Neural Networks

Autores: Pedro Jaquetti, Heitor Silvério Lopes

Resumo: Drive assist in agriculture has emerged as a promising solution for increasing resource efficiency and reducing operating costs. This paper details the development of essential components for an agricultural drive-assist system: a lightweight semantic segmentation model for crop row perception and a local path planning algorithm that extracts a navigation trajectory from the model’s output. The segmentation model employs a U-Net-style decoder with a MobileNetV3 encoder, optimized for computational efficiency. The system was rigorously evaluated on a custom dataset from orange groves and vineyards using 5-fold cross-validation. To confirm its suitability for real-world deployment, the segmentation model’s performance was benchmarked on an RK3588 System-on-Chip (SoC). Our approach achieved a mean Intersection over Union (IoU) of 94.76% (±0.24%), outperforming a DeepLabV3+ baseline in both accuracy and stability. On the RK3588, the model demonstrated an inference speed of 46.4 FPS. These results validate that the proposed components provide a robust, stable, and efficient foundation for vision-based drive-assisted on resource-constrained agricultural hardware.

Palavras-chave: Computer vision; Deep learning; Drive assisted; Vision-based autopilot.

Páginas: 6

Código DOI: 10.21528/CBIC2025-1190697

Artigo em PDF: CBIC_2025_paper1190697.pdf

Arquivo BibTeX:
CBIC_2025_1190697.bib