| With the continuous development of artificial intelligence,computer vision techniques have become increasingly important,and 2D human pose estimation,as a basic but very challenging computer vision task,has become a hot research topic in recent years.Early human pose estimation algorithms rely on hand-designed appearance features and graphical structures,which cannot fully extract high-level semantic features of images.Existing human pose estimation algorithms have made great progress by using deep convolutional neural networks for end-to-end keypoint prediction.However,the detection results are still different from expectations because of problems such as blurred appearance,difficulty in locating blocked keypoints,and difficulty in capturing the semantic relationships between human keypoints and the global relationships of the image by convolutional neural network models.Accurate human pose estimation algorithms need to address several key issues:first of all,the problem of how to better predict keypoints which are difficult to locate due to occlusion,blurred appearance,complex background,etc.;then,the problem that current convolutional neural network-based human pose estimation algorithms have difficulty capturing the global dependence of images and the semantic relationships between keypoints of the human body;and finally,the problem that most current multiscale fusion methods ignore the problem of cross-channel communication of features at different scales.To address the above key scientific issues,the main research contents and innovations of the paper are as follows.1.A new two-stage human pose refinement framework based on Heatmaps-guided Keypoints Encoder and Pose Refinement Module is proposed to address the problem of blurred appearance and blocked keypoints that are difficult to locate.The framework uses the commonly used heatmap model to predict the keypoint heatmaps in the first stage,and generates 1D keypoints embeddings using the heatmap guidance in the second stage,which makes the generated keypoints embeddings focus more on specific regions rather than the whole image range and reduces the fusion of redundant semantic information,after which the keypoints embeddings are fed into the pose refinement module to obtain refined prediction results.The experimental results show that the framework has the advantages of high accuracy,irrelevance of the backbone model and few additional parameters.2.To address the problem that convolutional neural networks are difficult to capture the global dependence of images and the semantic relationships among human keypoints,a new algorithm for human pose estimation,a Global-Local network based on Transformer and graph convolution is proposed.The algorithm proposes a new graph convolution PGConv(Pose Graph Convolution),which abandons the weight sharing scheme for a specific keypoint,and the distance from the keypoint is 1 as the short-range key point and the distance greater than 1 as the long-range keypoints.For the short-range keypoints and the long-range keypoints,different transformation matrices are used to extract the local information among the keypoints.In addition,the global dependence of the image is captured by the Transformer,and then the local and global information are aggregated to achieve more accurate human pose estimation.The experimental results show that the algorithm is able to obtain considerable detection performance.3.To address the problem that most current multiscale fusion methods ignore cross-channel communication of features at different scales,a human pose framework for multiscale fusion based on atrous convolution and channel shuffling is proposed.The framework obtains multi scale feature representations by cascaded atrous convolution,and then realizes cross-channel communication between different scale features with minimal parameter cost by channel shuffling operation.The feature is weighted by the attention module,and online hard keypoints mining(OHKM)is introduced to explicitly process the keypoints that are difficult to locate.Experimental results show that the algorithm can achieve considerable detection performance. |
