6D attitude estimation method

Abstract

The invention discloses a 6D attitude estimation method, and belongs to the field of attitude estimation, and the method comprises the steps: S1, obtaining a mask of a target on an image through image segmentation; s2, extracting a target image block and an XYZ map according to the mask range, and inputting the target image block and the XYZ map into the XYZNet for feature extraction to obtain dense 2D-3D features; s3, inputting the dense 2D-3D features into three CNN branches, and respectively regressing 3D translation, rotation quaternion and confidence to estimate the 6D pose of the target object; and S4, selecting a group of attitude results with the maximum confidence coefficient, storing point cloud data in an XYZ map format through the method and the XYZNet, enabling the point cloud to be strictly aligned with the image, and performing 2D-3D feature extraction on the RGB image and the XYZ map by using 2D convolution, so that the design utilizes a convolution window structure to aggregate points originally in a 3D space, and can efficiently extract local geometric features of the point cloud. The PrimADD (S) loss is provided, so that the algorithm can correctly process the symmetric object in the training stage, and meanwhile, the estimation of the asymmetric object is also improved.

Description

technical field [0001] The invention relates to the field of attitude estimation, in particular to a 6D attitude estimation method. Background technique [0002] The purpose of 6D pose estimation is to detect objects in the scene and estimate the rotation and translation of the object relative to the canonical frame. Efficient and accurate 6D pose estimation is a key technology for real-time interactive applications, such as augmented reality, autonomous driving, and robot applications. Designing a robust pose estimation algorithm is challenging due to changes in illumination, data noise, and incomplete occlusions in real scenes. In recent years, thanks to the powerful feature extraction and fitting capabilities of deep learning, data-driven pose estimation algorithms have shown excellent performance. These deep learning-based attitude estimation algorithms can be divided into 2D image-based methods, 3D point cloud-based methods, and 2D image-3D point cloud-based methods a...

AI Technical Summary

Problems solved by technology

[0003] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to propose a 6D pose estimation method, through which the problems of inefficient extraction of dense 2D-3D features and rotational ambiguity of symmetrical objects are solved

Images