<p>Human pose estimation is a key step for computers to understand human behavioral intentions. Although existing bottom-up regression-based aggregation methods are efficient, they often produce incorrect connections of joints between different instances in crowded or occluded scenes due to their inability to handle pose ambiguity and limited scale adaptability. To address these issues, this paper introduces keypoint density-guided scale-adaptive convolution to respond to spatial variations in crowd distribution. Specifically, a backbone network with a small receptive field is used to extract dense keypoint representations from the image, after which a density estimation algorithm based on the Keypoint Expansion Map (KEM) is designed to adaptively adjust the local receptive field window size according to the sparsity of the keypoint distribution. Empirical observations show that predictions of backbone networks with fixed receptive fields usually exhibit only small local deviations. Based on this, we propose a scale-enhanced pose fine-tuning module based on a transformer architecture. It utilizes enhanced representations with the correct receptive fields to perceive the offsets of initial pose predictions, thereby refining them. Meanwhile, it establishes long-range dependencies among sparse joints, effectively reducing incorrect connections. Finally, we perform model regularization at the inference stage to achieve a balance between accuracy and efficiency. Extensive evaluations on multiple datasets demonstrate that the proposed method significantly improves estimation accuracy in crowded scenes, which verifies its potential for real-world applications.</p>