The human pose estimation task aims to estimate the pose of human body from given images or videos. The maintain human pose estimation algorithms mainly focus on the prediction of keypoints, without considering the connection relationship between two adjacent keypoints, and the precise localization property of shallow features, which leads to sub-optimal solutions. To thoroughly investigate the significance of these two essential yet neglected factors, we propose a novel human pose estimation method named Linlot, composed of Limb connection relationship constraints and keypoint localization refinement module. Specifically, we devise a limb orientated (LO) loss to strengthen the connection relationships between adjacent keypoints, and design a lightweight plug-and-play keypoints localization refinement (KLR) module which leverages precise localization information of shallow features for refinement keypoints localization. Our Linlot effectively reduces bizarre poses, and improves the performance of occluded keypoints prediction. It should be noted that our proposed LO loss and the KLR module are effective even when used alone. Besides, when combining these two components, our Linlot achieves satisfactory performance in human pose estimation. Extensive experimental results on two mainstream benchmark datasets, COCO and MPII dataset, demonstrate the effectiveness of our proposed Linlot. Code is available at https://github.com/Loitertwx/Linlot .