Reasoning-based Multi-level Predictions For Single-person Human Pose Estimation

With the rapid development of technology and the fast updates on social media,it is the key for the development of Internet intelligence to process increasingly massive image and video data by human behavior analysis technology.Human behavior analysis is one research focus in the field of computer vision.Single-person human pose estimation is the basis of various human behavior analysis techniques,which is essential for image and video understanding.The human keypoints are usually small and easy to be occluded,making it difficult for the design of highly accurate algorithms.Nowadays,more and more complex convolutional neural networks(CNN)are designed for improving the accuracy of keypoints.However,without deeply mining the level of detection difficulty for keypoints,the increase of their accuracy is extremely small and easy to appear inconsistent postures,so it is difficult to meet the requirements for the high accuracy of human pose estimation in practical applications.In this paper,reasoning-based multi-level predictions for single-person human pose estimation is proposed to obtain the high accuracy of human keypoints detection.The main work of this paper includes the following two aspects:Firstly,in order to solve the imbalance of feature learning for different keypoints,a multilevel prediction based single-person pose estimation network is proposed.This method subdivides keypoints into three different levels according to their detection difficulty,and uses different cascaded sub-networks to predict their location accordingly.During the training phase,different-level samples can be fully trained,so that the model learns more discriminative feature representations.Experimental results show that this method can greatly improve the detection accuracy of difficult keypoints,while ensuring the high detection precision of the easy examples.Thereby,the detection performance of the entire model has been greatly improved.Secondly,for the problem of inconsistent poses cased by the interference noise in the detection heatmaps,a graphic model reasoning based single-person pose estimation network is proposed.By means of the rich spatial and semantic information in heatmaps,this method explicitly constructs the structure information of human pose and graphic model reasoning network can implement the informative interactions between heatmaps.On the one hand,this network can add useful background information to the keypoints on the semantic level,which can enhance the expression of heatmaps.On the other hand,this network can explicitly make use of the consistency constraints of keypoints to suppress the interference noise in heatmaps,which enhances the discriminative representation of keypoints features.The experimental results show that this method can well solve the problem of inconsistent pose estimation caused by the interference noise in heatmaps,which can achieve highly accurate detection of human targets in images.In summary,reasoning-based multi-level predictions single-person human pose estimation can achieve highly accurate pose estimation results in humans in natural images,surpassing the existing single-person pose estimation methods,which has certain theoretical research value and practical application value.