In recent years, RFID indoor localization has emerged as a promising technique for precise spatial tracking within complex indoor environments. However, traditional positioning algorithms often struggle with limited computational resources, noisy data, and the dynamic nature of indoor layouts. Machine learning has emerged as a critical enabler for overcoming these limitations and improving the accuracy and reliability of RFID-based localization systems. Deep learning techniques, in particular, have shown significant potential by leveraging large datasets and complex feature representations to enhance positioning performance. Despite these advancements, existing approaches still face challenges, including limited precision in coordinate prediction, reliance on insufficient training data, and so on. To address these issues, we propose a novel model that integrates bidirectional dilated gated recurrent units (BiDGRU) to capture long-term dependency patterns within sequential data, followed by a sparse attention mechanism (ASSA) to refine feature representations. Finally, the Transformer architecture is employed to encode and decode spatial information, enabling accurate coordinate prediction. The experimental outcomes demonstrate that the proposed method markedly enhances positioning precision. It not only overcomes the inherent limitations of traditional algorithms but also paves the way for scalable solutions in dynamic indoor environments. With the continuous proliferation of Internet of Things (IoT) devices and the increasingly widespread application of 5G technologies, such an RFID indoor positioning system is likely to play a role in fields such as smart retail, industrial automation, and smart cities.