<p>PIDNet provides a favorable accuracy–speed balance for real-time semantic segmentation, but its original single-modality design is limited when directly applied to RGB-T scenes, where visible and thermal features show heterogeneous responses and require stronger detail preservation and semantic collaboration. To address this issue, this paper proposes DGHF-PIDNet, a detail-preserving and hierarchically guided PIDNet framework for real-time RGB-T semantic segmentation. The proposed method retains the original P–I–D parallel structure and redesigns branch interaction from two aspects. First, Dynamic Asymmetric Gated Re-parameterizable Convolution (DAGRConv) is inserted into the high-resolution P branch. During training, it uses asymmetric multi-branch transformation and input-dependent channel-wise gating to enhance local structures, directional boundaries, and fine-grained details; during inference, its branches are re-parameterized into a single 3 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> 3 convolution for efficient deployment. Second, Hierarchical Fusion Bidirectional Gate (HF-BIGate) replaces the original P–I fusion units and uses compressed semantic features from the I branch to progressively guide and adaptively update the P branch through bidirectional gated interaction and attention-guided fusion. In this way, DGHF-PIDNet strengthens detail representation, semantic guidance, and boundary consistency within a compact real-time architecture. Experiments on SemanticRT, FMB, and PST900 show that DGHF-PIDNet consistently improves over the PIDNet baseline while maintaining 7.85M parameters and 218 FPS, demonstrating a favorable balance between RGB-T segmentation accuracy and real-time efficiency.</p>