| The 2D human pose estimation task is mainly solved by deep learning methods and has made great progress in recognition accuracy with the development of deep learning.At present,the existing 2D multi-person pose estimation algorithms are mainly two-stage algorithms,which are not trained end-to-end.At the same time,the requirements for computing power are high,which limit the deployment and application of the algorithm on embedded platforms.Aiming at the shortcomings of the current two-stage multi-person pose estimation algorithm strategy,based on the cutting-edge work,combined with the requirements of real-time and rapid deployment of the algorithm on the embedded platform,this thesis further studies the 2D multi-person pose estimation algorithm.The specific work is as follows:1.This thesis chooses one-stage multi-person pose estimation algorithm as the baseline route of research and designs the multi-person pose estimation baseline algorithm framework based on this route.Then,three different backbone networks Res Net-50,DLA-34 and HRNet-w32 are used to train and test on the COCO dataset and achieve good accuracy,which verifies the feasibility of the baseline algorithm framework.2.Aiming at the serious deviation between the receptive field of the output feature of the head2 branch of the baseline framework and the actual position of the human keypoints to be predicted,a module named Feature Alignment Module based on Body Parts(FAMBP)is proposed,which dynamically uses the feature vector of the human part closer to the human keypoints for further prediction during the model learning process.FAMBP achieves an average m AP improvement of 8.53% on COCO dataset while only introducing a small amount of inference delay.3.The head2 branch in the baseline algorithm framework directly uses the L1 loss function,assuming that the keypoint labels conform to the Laplacian prior distribution,which is simple but not optimal.In this thesis,the residual log-likelihood estimation(RLE)is introduced into the multi-person pose estimation method.The normalizing flow model is used to learn the distribution of the error between the human keypoint labels and the predicted value,so as to estimate the distribution probability density of the human keypoints,design a loss function that is more in line with the distribution of keypoints,and improve the supervision effect during training.RLE achieves an average m AP improvement of 2.17% on the COCO dataset using three backbone networks without additional introduction of inference delay.4.This thesis makes an in-depth study on the deployment and optimization of the multi-person pose estimation algorithm proposed in this thesis on embedded platforms.The algorithm is deployed on two actual embedded platforms: Nvidia Jetson Xavier NX and Tiangong AT210,and the model quantization technology is used to greatly reduce the demand for memory and computing resources of the algorithm model.The inference speed is greatly accelerated while the accuracy of the algorithm only decreases slightly.Based on the one-stage multi-person pose estimation algorithm strategy,this thesis proposes an multi-person pose estimation algorithm suitable for deployment and application on the embedded platform,and makes improvements and optimizations for the shortcomings of the algorithm.The detection effect of the algorithm is improved without affecting the inference speed of the algorithm.The algorithm has certain practicability and application value. |
