Indoor attitude estimation method based on thermodynamic map for single image

Abstract

The invention belongs to the field of target attitude estimation, and discloses a single image indoor object attitude estimation method based on thermodynamic map. The object attitude estimation method based on thermodynamic map extracts candidate frames of a plurality of target objects through an RPN network. The thermodynamic images of eight vertices of each object bounding box on two-dimensional images are extracted by full convolution network (FCN), and then the 6D attitude estimation of the object is calculated by using PnP method. ShapeNet is used as CAD model library to synthesize a large amount of training data. The object posture estimation detection technology based on the thermodynamic map adopted by the invention has strong robustness, can estimate the posture of different indoor objects when the background is disorderly indoor scene and the object is partially occluded, has wide application range, is insensitive to light and does not require the object to have obvious texture appearance.

Description

technical field [0001] The present invention relates to the technical field of pose estimation, in particular to a method for estimating the pose of an indoor object in a single image based on a thermal map. Background technique [0002] Pose estimation of objects in indoor scenes plays an important role in the motion planning of social robots and human-computer interaction in virtual reality and augmented reality. At present, in the research of pose estimation, it is mainly divided into feature point matching, template matching, dense feature matching and end-to-end methods based on convolutional networks. There are certain problems in these methods, and the performance is not very stable in complex practical environment applications. For example, the traditional method based on feature point matching relies on texture to extract feature points, and then calculates the rotation and translation between the corresponding points of the 3D object based on the 2D feature points...

AI Technical Summary

Problems solved by technology

[0003] To sum up, the problems existing in the existing technology are: the traditional method based on feature point matching cannot deal with smooth and textureless objects; the method based on template matching is sensitive to illumination and occlusion; the method based on dense feature matching needs to The sample space is used to extract features, which is time-consuming and the posture generally requires subsequent optimization; the end-to-end method based on the convolutional network is not good at solving multi-target and complex scenes and occlusion between objects, and cannot meet the application requirements of higher accuracy

Images