Abstract: Irregularly shaped small fields are widely distributed across hilly agricultural regions, where curved and undulating headland boundaries pose persistent challenges for autonomous navigation. Achieving unmanned edging operations in such environments required robust perception of complex field structures and the ability to generate smooth, safe obstacle avoidance paths while maintaining unilateral guidance along the inner headland boundary. Traditional machine vision techniques were often inadequate for this task due to low robustness under varying field conditions, limited spatial understanding, and difficulty distinguishing subtle texture differences between headland and farmland. Likewise, conventional local path planning methods were not well suited for unilateral guidance scenarios, as they typically failed to ensure both a smooth obstacle-bypassing process and a stable, consistent rejoining of the reference boundary. To address these challenges, this paper proposed a unilateral guidance obstacle avoidance path planning method based on RGB-D multimodal data, which consisted of two key components: high-fidelity scene perception map construction and intelligent real-time path planning. At the perception level, a depth camera simultaneously captured color and depth information. An efficient RGB-D semantic segmentation model was adopted to exploit the complementary strengths of RGB texture cues and depth geometry. The network followed a dual-stream encoder–decoder design with multilevel cross-modality fusion, attention-based feature reweighting, and multi-scale supervision, which enabled accurate pixel-level classification of headland and obstacle regions under diverse lighting, crop residue, and soil conditions. Depth information enhanced separability in cases where RGB contrast was weak, such as dry field headland edges, thereby ensuring stable segmentation performance across different terrains. To transform segmentation into spatially meaningful representations, a three-stage point cloud processing pipeline was developed. First, each RGB pixel was back-projected into 3D space using depth measurements and camera calibration parameters to form a dense point cloud. Second, semantic labels from the RGB-D segmentation were transferred to the corresponding 3D points through pixel-point indexing, resulting in a semantically annotated point cloud. Third, a ground-plane estimation and projection procedure generated a 2D obstacle-avoidance coordinate map by projecting all relevant points onto a common field plane. The resulting coordinate map provided a compact yet informative representation that is well suited for efficient, real-time path planning. On this basis, an enhanced artificial potential field algorithm was designed to meet the unique demands of single-edge-guided obstacle avoidance. Several key improvements were introduced, including geometric generalization of repulsive fields to handle arbitrarily shaped obstacles, region-aware gain modulation to regulate repulsive forces near the headland boundary, integration of a smoothing factor to eliminate curvature discontinuities, headland-specific repulsion modelling to maintain a stable lateral offset from the boundary, and sub-target generation to avoid local minima caused by force equilibrium. The resulting planner generated obstacle avoidance trajectories with gentle obstacle approach, early exit, and overall path smoothness, while maintaining the desired distance from the headland. To facilitate direct use by vision-based path tracking controllers, the planned trajectory was further re-projected onto the forward-looking RGB image through inverse coordinate transformation, so that the downstream control module received visually aligned path references. Extensive experiments were conducted to validate the effectiveness of the proposed framework. The RGB-D segmentation network achieved a highest mean Intersection over Union of 95.97% on the paddy-field and dry-field datasets, consistently producing accurate and stable segmentation results. The improved potential field planner generated obstacle avoidance trajectories with path deviation below 9 pixels and standard deviations under 1.2 pixels across various obstacle shapes and placements. Furthermore, real-world field experiments demonstrated that the system achieved an average lateral deviation below 0.076 m during autonomous operation. The overall processing efficiency reached 23.19 frames per second, indicating that the system satisfied real-time deployment on agricultural machinery. Overall, the proposed framework provided an integrated solution for autonomous obstacle avoidance under unilateral headland guidance. By combining multimodal perception, 3D spatial reasoning, and behavior-aware planning, the proposed method enabled reliable navigation in irregularly shaped small fields and contributes a practical approach toward fully unmanned agricultural operations.