Multi-person Human Pose Estimation Based on Deformable Convolution

Deep neural networks for human pose estimation all sample at the fixed position of the feature map, and therefore it is difficult to model the geometric transformation of human pose. The generalization ability of the network is poor with the variation of the size, pose and shooting angle of the human instance. To solve this problem, multi-person human pose estimation based on deformable convolution is proposed.Based on the strong ability of deformable convolution in modeling geometric transformation of targets, a feature extraction module is designed to ensure the detection accuracy under the geometric changes of human key points. To further improve the performance of the network, the prediction value of the model and the truth value generated by the two-dimensional Gaussian model are employed to calculate the loss, and the model is trained iteratively. The human key points are detected effectively by the proposed model under the complex conditions, such as shooting angle, attachment and character scale changes. The experiment shows that the proposed model effectively improves the accuracy of human key point detection. © 2020, Science Press. All right reserved.