Scope: Occlusion among pedestrians is very common in crowds, which greatly impairs pose estimation and subsequent pedestrian behavior understanding. This task mainly aims to improve the performance of the pose estimation model in crowded scenarios by constructing a interaction based occlusion-aware pose estimation model. Objectives: - Optimize AlphaPose’s detection post-processing process by introducing an adaptive NMS strategy, reducing the overlap and missing rates of human detection boxes in crowded scenes. - Design a spatial occluded keypoint inference module, combining skeleton priors and interaction information to improve the detection accuracy of fully occluded keypoints. - Improve the multi-scale feature fusion mechanism to enhance the model’s adaptability for pose estimation of small-scale humans in crowded scenes. - Verify the pose estimation AP&AR performance of the optimized model on public datasets and real-world scenes.