Human pose estimation is a fundamental task in pattern recognition and intelligent robotics, with broad applications in areas such as human-computer interaction and autonomous systems. Despite recent advances, existing Transformer-based methods often suffer from high computational cost and degraded performance under occlusion. To address these issues, we propose an efficient Boundary-Aware Former (BAFormer) block that enhances long-range dependency modeling via adaptive window boundary constraints and incorporates convolutional operations to promote interaction and aggregation among features across different windows. This design effectively reduces model complexity while improving keypoint localization accuracy, especially under occluded conditions. By stacking multiple BAFormer blocks, we introduce BATPose, a Boundary-Aware Transformer for 2D human pose estimation, available in two variants: BATPose-B and BATPose-L. Additionally, a SimMIM-based pre-training strategy tailored for BATPose is adopted to further enhance its performance. Extensive experiments on four widely used benchmark datasets demonstrate that BATPose achieves competitive performance with significantly reduced computational cost (23%–45%), reaching 77.2 AP on the COCO val2017 dataset and maintains strong robustness under occlusions. Compared to recent state-of-the-art methods, BATPose delivers a compelling balance between accuracy and efficiency.