Stacked object 6D pose estimation method and device based on deep learning

Abstract

The invention discloses a stacked object 6D pose estimation method and device based on deep learning. The method comprises the steps: inputting a point cloud of scene depth information into a point cloud deep learning network, and extracting the features of the point cloud; learning semantic information of an object to which point cloud features belong, a foreground and a background of a scene and3D translation information of the object to which the point cloud features belong through a multi-layer perceptron, and performing regression to obtain seed points; randomly sampling K points in theseed points, K being greater than the number of to-be-estimated objects, and classifying the seed points by taking the K points as central points; predicting 3D translation, 3D rotation and 6D pose confidence of the object according to the features of each type of points through a multilayer perceptron; According to the predicted 6D pose and the 6D pose confidence coefficient, using a non-maximumsuppression NMS method to obtain a final scene object pose. According to the method, accurate estimation of the pose of the stacked scene is realized end to end, the input is the scene point cloud, the pose of each object in the scene is directly output, and the shielding problem of the stacked objects can be well solved.

Description

technical field [0001] The invention relates to a method and device for estimating the 6D pose of a stacked object based on deep learning. Background technique [0002] Bin picking is a core problem in robotics and computer vision. The goal is to allow a robot with a visual sensor to grab randomly placed objects through an end effector (suction cup, two-finger manipulator). In order to accurately grasp and place objects, 6D pose estimation of objects is required, that is, rotation and translation. The translation is the XYZ coordinates of the object, and the rotation is the attitude of the object, which can be represented by a rotation matrix (or Euler angle, or unit four elements). [0003] In the case of serious stacking, it is difficult to estimate the pose of all stacked objects. Traditional methods use local feature matching in RGB images, such as SSD-6D ([1] W. Kehl, F. Manhardt, F. Tombari, S. Ilic, and N. Navab, "Ssd-6d: Making rgb-based 3d detection and 6d pose e...

AI Technical Summary

Problems solved by technology

The disadvantage is: PPR-Net needs to be clustered after the network output, and clustering needs to set some hyperparameters, and different parameters need to be set for different objects, which makes the algorithm framework still rely on manual debugging in the later stage

At the same time, point-by-point regression consumes a lot of computing time and computer memory in the case of a large number of scene points

Images