EFNetPX: An improved panoptic perception network for inter-row navigation in orchard

• Proposing a novel multi-task network EFNetPX model integrating drivable area segmentation and obstacle detection tasks. • Improvement of the segmentation head compared to pure semantic segmentation networks by introducing a trunk detection task. • Developing a navigation method with reference line fitting and driving safety assessment based on panoptic perception. Drivable area segmentation and obstacle detection are critical for inter-row navigation in orchards. However, existing methods typically address these tasks independently, leading to increased computational requirements. In this study, we proposed a novel panoptic perception network, EFNetPX, that integrates drivable area segmentation and obstacle detection tasks. First, we constructed an orchard dataset containing 1335 images featuring various fruit trees and obstacles. Additionally, based on YOLOPX, we enhanced the backbone with the improved EfficientFormerV2, introduced a BiFPN neck, and applied Powerful-IoU loss function. EFNetPX’s anchor-free design improves generalization across orchard scenes. Finally, the RANSAC algorithm was applied to extract the navigation line based on the semantic segmentation results, combined with the obstacle information to assess driving safety. EFNetPX improved CPU inference speed from 4.93 frames per second (FPS) to 8.03 FPS, while obtaining a mean average precision (mAP) of 93.92 % and a mean intersection over union (mIoU) of 97.21 %, with results comparable to the baseline model (93.37 % mAP, 97.26 % mIoU). Notably, introducing the trunk detection task improved the mIoU value by approximately 7 percentage points compared to pure semantic segmentation networks such as STDC. The attitude error and distance error of line extraction on 640 × 360 resolution images were within 1.19 degrees and 6.96 pixels, respectively—representing reductions of 41.82 % and 39.24 % compared to results based on trunk detection. This improvement significantly enhanced navigation robustness in complex environments. The proposed method could provide an effective environment-aware solution for agricultural robotics applications.