Multi-person pose estimation in complex scenarios faces a critical challenge in balancing detection accuracy and computational efficiency. While top-down approaches incur prohibitive computational costs with increasing human instances, bottom-up methods suffer from keypoint grouping errors under dense crowds. To address this accuracy-efficiency dilemma, we propose VSS-Pose, a novel single-stage framework built upon the YOLO architecture. This framework integrates the CBS-VSS structure, which enhances keypoint localization accuracy through the Selective State Space Mechanism and Cross-Scan Module. It also employs the Dysample upsampling technique to optimize feature capture for complex poses or occluded keypoints and incorporates the Adown module to reduce model parameter complexity while maintaining detection accuracy. Experimental results on the CrowdPose benchmark dataset demonstrate that our largest model, VSS-Pose-L, achieves an AP of 74.2% on the CrowdPose dataset, outperforming state-of-the-art single-stage pose estimation models with only 38.0 M model parameters.