Understanding and leveraging scene semantics is critical for efficient object navigation (ObjectNav). However, existing end-to-end frameworks often struggle with limited interpretability and inadequate semantic reasoning. In this paper, we introduce SMO-RG, a modular target prediction method that fuses semantic maps and object relation graphs to enhance scene understanding and navigation performance. By capturing both global spatial context and local object relationships, SMORG enables more accurate prediction of target locations. We integrate SMORG into a modular ObjectNav pipeline and validate its effectiveness through extensive experiments on the HM3D simulation dataset and in real-world environments. Results demonstrate that SMORG significantly improves navigation success rates and predictive accuracy, highlighting its potential for robust ObjectNav in complex, unseen settings.