Research And Implementation Of Human Pose Estimation Algorithm Based On Deep Learning

In recent years,due to the wide application in security monitoring and sports guidance,human pose estimation has gradually become a popular topic in computer vision,meanwhile the rapid growth of market demand has also brought challenges to human pose estimation technology.The existing deep learning-based methods essentially use a single mapping function to lift the joints of the 2D human pose to the 3D space.When dealing with an uneven multi-distribution dataset,the weight of outlier samples will be reduced in order to ensure the overall accuracy.Specifically,these outlier samples refer to large-pose samples,which are common in real-life scenes.This causes a huge precision loss when existing methods migrating to natural scenes.This paper proposes a multi-branch lifting network to focus on the prediction of large-pose samples:firstly,we obtained the rough correspondence between the pose and human center of gravity by mining the internal connections of 2D joints,so that we are able to realize unsupervised pose classification by clustering the center of gravity using the K-means algorithm,after that the large pose samples and the normal ones are separated.Then multiple weak regressors are constructed to deal with different types of poses,so that the solution of human pose estimation changes from single-map fitting multi-distribution data to multi-map fitting multi-distribution data.The workload of each weak regressor is reduced,making it focus more on the prediction of one specific pose;Finally,drawing on the idea of integrated learning,multiple weak regressors are fused into a strong regressor,which is used to predict the pose from all distribution.The main contributions of our work are as follows:1.By mining the internal connections between 2D joints,we find the rough correspondence between human pose and center of gravity,and then use the K-means clustering algorithm to achieve the unsupervised classification of poses through the center of gravity clustering.2.Different from the previous methods,we design a multi-branch lifting network to handle the poses with different distributions,where each branch acts as a weak regressor and which is responsible for the prediction of one specific category of poses.Thanks to unsupervised pose classification,we can achieve specific training by modifying the weights of the dataset.3.We design different initial 3D poses for each branch according to its tasks for efficient training,at the same time limiting the scope of network exploration.The optimize target of the network changes from predicting 3D poses to optimizing the initial 3D poses.The task of each branch is further refined,the training efficiency and prediction accuracy were improved.4.We design and implement a 3D human pose estimation display system.The system introduces the data processing and algorithm flow of this paper,and which can perform 3D human pose estimation with 2D human joints sequences.