<p>To address challenges in detecting small objects on tactile paving, such as weak features, complex occlusions, and dense interference, a lightweight real-time detection algorithm called GDE-YOLO is proposed. Firstly, the C2f with Gated Bottleneck Convolution (C2f-GB) module replaces the traditional convolution with gated bottleneck convolution, and combines low-rank projection, depth separable convolution, and pixel-level gating strategies to significantly enhance the edge details and feature robustness of obstacles while compressing parameters and reducing computational costs. Second, we propose the DySample module, which achieves pixel level feature recovery through content-aware shifting. This approach halves computational complexity while enhancing localization accuracy for small and occluded objects. Finally, an enhanced feature aggregation head (EFAH) with decoupled detection is constructed, embedding a cross-layer local attention mechanism to adaptively fuse multi-scale features, significantly enhancing the model’s positioning ability for small targets. Experimental results show that this model achieves an accuracy of 69.9% and a FPS of 106.39 while significantly reducing volume and inference delay, demonstrating excellent real-time detection performance and meeting the requirements for real-time detection of tactile paving obstacles.</p>