Abstract: Vision-based autonomous navigation of agricultural machinery is often confined by variable natural illumination, irregular vegetation shadows, weed interference, and heterogeneous path morphologies under complex environments in orchards. These visual interferences can cause discontinuous navigation lines, leading to low accuracy of path recognition. Furthermore, the computationally intensive segmentation models are not expected to fully meet the requirements of real-time operation. It is required to balance recognition accuracy and real-time efficiency for navigation-line continuity. In this study, a robust path-recognition and navigation-line-extraction framework was developed using an improved U-Net model and a dynamic region-constrained bimodal centerline extraction (DRBCE). An improved U-Net semantic segmentation model was constructed to integrate a convolutional block attention module (CBAM), a multiscale feature-fusion mechanism, and deep supervision. The DRBCE algorithm was then designed to treat discontinuous segmented regions. The images were associated with either two visible row boundaries or only one visible boundary. Smooth navigation lines were generated after statistical filtering and cubic B-spline interpolation. A total of 1 630 images were collected from standardized orange and pomelo orchards, a citrus breeding orchard, and a tea plantation. The datasets were divided into training, validation, and independent test sets at a ratio of 7:2:1. A series of tests were performed, such as segmentation, challenging-scene evaluation, navigation-line extraction, edge-cloud deployment, and 15-m field tracking. On the independent test set of 163 images, the improved model achieved a mean intersection over union (mIoU) of 97.99%, an accuracy of 99.01%, a precision of 98.94%, a recall of 99.04%, and an F1-score of 98.98%. Its single-image inference time was 0.10 s. The final configuration contained 18.86 million parameters. Compared with the basic U-Net, the improved model increased mIoU and accuracy by 1.52 and 1.27 percentage points, respectively, while the inference time was reduced from 0.12 to 0.10 s. Compared with the basic U-Net, DeepLabV3+, and E-Net models, the highest accuracy of segmentation was obtained to retain practical efficiency. Ablation tests showed that CBAM, multiscale feature fusion, and deep supervision contributed most to feature enhancement, spatial recovery, and training optimization. In the three representative challenging-scene subsets, with 50 images in each subset, the mIoU values of 82.74%, 99.15%, and 97.97% were obtained under weed-covered, bare-ground, and complex-boundary conditions, respectively. The mIoU values were only 10.61%, 45.06%, and 40.78% using the Excess Green with Otsu thresholding. In navigation-line extraction, compared with the conventional A-star algorithm, the average processing time was 0.68 s and the path-length error was 32.11 pixels, representing reductions of 45.16% and 45.28%, respectively. Its maximum central processing unit (CPU) utilization was 73.53%, which was 17.58% lower than that of the A-star algorithm. The standard deviation of path curvature decreased from 0.2182 to 0.1014, indicating a 53.5% reduction in curvature variability. After deployment in an edge-cloud collaborative architecture, the Raspberry Pi 4B edge node processed images at an average rate of 6.47 frames per second (FPS), with average CPU and memory utilization rates of 33.88% and 30.37%, respectively, and an average command latency of 5.97 ms. The average single-frame inference time on the cloud node was 0.04 s. In field tracking at 0.3 m/s, a 15° steering deadband threshold was smoothly completed the 15-m route, with a mean steering deviation below 4.80°. The improved U-Net and DRBCE framework accurately recognized traversable orchard paths to generate the smooth, continuous navigation lines under the visual interference. The edge cloud fully met the real-time requirements of low-speed fields, while the optimal steering deadband improved the tracking stability of the mobile platform. Therefore, it is expected to integrate semantic segmentation, centerline extraction, and steering control into vision navigation under additional terrain, weather, and vehicle conditions. The findings can provide a technical reference for autonomous navigation of orchard machinery in different inter-row scenes.