A Single-Stage Navigation Path Extraction Network for agricultural robots in orchards

Abstract

The real-time and precise extraction of navigation paths holds significant importance in ensuring the autonomous navigation of agricultural robots. Although widely used in orchards, path extraction for agricultural robots remains a complex, multi-stage process. To address the limitations of current vision-based algorithms, this paper proposes a novel approach: the Single-Stage Navigation Path Extraction Network (NPENet). NPENet simplifies the path extraction process by reducing unnecessary parameterization and redefining the road centerline as the neural network’s primary prediction target, with a corresponding tailored loss function. Utilizing residual modules, NPENet effectively extracts navigation path features in orchard environments. The model’s performance is further enhanced by optimizing the network structure. A dataset of 25,720 images from various orchard scenes was used to train and test the model. Experimental results demonstrate that NPENet achieves 92.14% accuracy in road centerline detection and 91.6% recall, with a detection speed of 10.1 ms per 448x448 pixel frame on a Jetson Xavier, and a parameter size of only 1.5 M. These findings show that NPENet outperforms existing visual detection and segmentation methods, providing efficient and accurate road information for mobile robots in orchard environments. This approach offers a promising solution for autonomous navigation in agriculture.