This study focuses on developing an end-to-end autonomous driving model that considers the contextual interplay between pixels, and proposes the SKGE-Swin architecture. SKGE-Swin utilizes the Swin Transformer, which leverages a skip-stage mechanism, to expand feature representations across multiple network levels and globally. Leveraging the Swin Transformer's Shifted Window-based Multi-head Self-Attention (SW-MSA) mechanism, it extracts information from distant pixels and retains important information from the initial to final stages, enhancing the ability to understand complex patterns in the surrounding environment. Using adversarial scenarios on the CARLA platform, we simulated and evaluated real-world environments, achieving driving scores superior to existing methods. Furthermore, we conduct an ablation study to evaluate the contributions of each architectural component, including the impact of skip connections and the use of the Swin Transformer.
