Accurately detecting farmland boundary lines is crucial for performing cross-row operations with autonomous agricultural vehicles. In dryland farming environments, complex and variable conditions, including weed interference, lighting variations, and difficult-to-identify headland area field ridges pose significant challenges for accurately detecting headland boundaries. Although the existing detection methods demonstrate good performance across various scenarios, considering the complex variations found in dryland farming environments could improve them. Therefore, an efficient and accurate DeepLabv3+-based headland boundary detection algorithm is developed. The method employs a lightweight MobileNet architecture, using depth information as auxiliary features to better perceive terrain variations. A multihead dual-feature attention mechanism (M-DAFM) adaptively fuses red-green-blue (RGB) and depth information. An improved dense residual atrous spatial pyramid pooling (DR-ASPP) structure provides enhanced feature extraction capabilities while maintaining computational efficiency. Multistage processing-based fusion and polynomial fitting are subsequently used to extract boundary lines from headland contours, and multiorder Bezier curves integrated with boundary curvature variation points are used to calculate autonomous turning paths for agricultural robots. Experimental results demonstrate that the improved DeepLabv3+ model achieves a mean intersection-over-union (mIoU) value of 93.9 % in headland boundary detection tasks with an inference speed of 23.1 frames per second (FPS). Final navigation experiments confirm that the robot can successfully perform autonomous cross-row turning operations with a 0.09- meter maximum deviation, demonstrating precise navigation control in complex agricultural environments. The proposed approach advances the theoretical agricultural robot navigation framework and provides a practical solution for implementing challenging autonomous cross-row operations in field conditions.