A novel precision navigation line extraction algorithm for apple orchards in multi-period complex scenarios

Accurately extracting inter-row navigation lines is a key technology for enabling vision-based orchard navigation in intelligent agricultural machinery. However, the extraction process is often affected by visual disturbances and non-standard tree row arrangements, leading to trunk-row fitting deviations and reduced navigation line accuracy. To address this, a navigation line extraction method integrating tree trunk detection is proposed. First, the YOLOv8 model is enhanced by replacing the standard convolutional layers with small-object detection layers and substituting the original detection head with the DCNv4-DyHead dynamic head. This modification improves the ability to detect small-sized objects and enhances adaptability to multi-scale targets. The experimental findings indicate that the improved model achieves a 4.98% increase in detection accuracy compared to the base model. Second, to mitigate fitting deviations caused by visual interferences such as tree trunks in adjacent rows, weed occlusion, and one-sided missing trunk, an Adaptive Trunk-Row Fitting Method (ATRFM) is developed. Experimental evaluations demonstrate that the line fitting accuracy of ATRFM reaches 93.33%, 96.67%, and 83.33% under the respective conditions of adjacent-row interference, weed occlusion, and one-sided missing trunk. Finally, by incorporating the perspective geometry characteristics of images to establish topological relationships, a Perspective-Guided Line Extraction (PGLE) method is introduced. Results indicate that in one-sided missing trunk scenarios, PGLE achieves comparable mean distance errors to traditional mean extraction methods. Moreover, in adjacent-row interference and weed occlusion scenarios, PGLE improves navigation line extraction accuracy by 16.67% and 26.66%, respectively, with mean distance errors of 1.28 cm and 2.67 cm. A novel navigation line extraction approach tailored to complex orchard environments is proposed in this work, offering technical foundations for the autonomous operation of intelligent agricultural machinery.