3D point cloud head pose estimation system and method based on ordered regression and soft labels

Abstract

The invention discloses a three-dimensional point cloud head attitude estimation system and method based on ordered regression and soft labels. The system includes: a feature learning network module for performing hierarchical feature extraction on point cloud data; a prediction network module for The feature obtained by the feature learning network module is mapped to the head attitude angle to obtain the angle prediction value, and the angle prediction value and the head attitude angle as the label are brought into the first loss function; the sorting network module is used to dimensionally divide the head attitude angle Form a number of subtasks, generate a soft label for the relationship between the head attitude angle as a label and the subtask, perform value prediction on the features obtained by the feature learning network module, and predict the value of the point cloud data and the soft label. Bring in the second loss function; network update module. The invention combines the loss of the sorting network module and the prediction network module, and its purpose is to introduce the sorting network, so as to guide the learning of the prediction network, make the feature extraction more accurate, and improve the accuracy of the prediction network.

Description

technical field [0001] The invention relates to the field of head pose estimation, in particular to a three-dimensional point cloud head pose estimation system and method based on ordered regression and soft labels. Background technique [0002] Robust head pose estimation is the basis of many problems in computer vision and computer graphics, and has a wide range of applications in human-computer interaction, such as: VR / AR, driver behavior analysis, etc. In recent years, head pose estimation using RGB images has made great progress and has been successfully applied to constrained scenes. However, challenges posed by unconstrained scenes, such as illumination changes, larger pose changes, and more severe occlusions, make RGB-based methods unreliable. Furthermore, the use of RGB images to convey facial appearance information is considered a threat to personal privacy, which is one of the major concerns of modern society and limits its further applications. [0003] With th...

AI Technical Summary

Problems solved by technology

[0007] The purpose of the present invention is to overcome the deficiencies of the prior art, provide a 3D point cloud head pose estimation system and method based on ordered regression and soft labels, and solve the problem that the hard regression data recognition effect of the prior art is not good

Images