Improved Yolo11 for Efficient Obstacle Detection in Navigation Systems

Current navigation assistance systems for the visually impaired often face a trade-off between obstacle detection accuracy and model efficiency. This paper proposes an improved detection approach based on the YOLO11 model to address this challenge. A dedicated high-quality data set of common road obstacles was first constructed that covers both dynamic and static types. The model architecture was then optimized to enhance feature extraction while reducing computational overhead. Specifically, we incorporate crossstage partial network design into the original YOLO11 by integrating partial convolution into the C3K2 module, effectively reducing redundancy while retaining feature richness. A lightweight Slim-Neck structure is further introduced to improve multi-scale feature fusion in the neck. To compensate for potential accuracy loss from lightweight components, a dynamic detection head is adopted to increase adaptability to complex targets. The experimental results demonstrate that the proposed method significantly reduces the number of parameters and computational complexity while achieving consistent improvements in detection accuracy. Ablation studies validate the contribution of each component. In general, the enhanced YOLO11 model achieves an effective balance between accuracy and efficiency, offering a promising solution for real-time road obstacle detection in navigation systems, particularly on resource-constrained mobile platforms.