利用RGB图像和DNN进行物体6DOf位姿推算

The Estimation of Object Position and Orientation Using the RGB Images and the DNN

提出一种基于RGB图片信息进行物体位姿推算的方法。方法利用二维图像分割得到的二值mask图像,并在此基础上进行物体的位姿推算,实现实时准确的物体6Dof位姿信息输出。目前传统上基于RGB-D图像推算物体的6Dof位姿精度较高,但RGB-D双目摄像头往往功耗和体积较大且成本高,而利用RGB图像来进行位姿推算无论在工业还是移动设备上更为实用。针对深度神经网络训练需要大量数据且数据集获取困难的问题,利用blender软件实现了一套物体6Dof位姿训练数据集生成的方法。通用实验表明,所提方法的推算的位姿精度相对较高,位姿推算用时约为0.1s,具有良好的准确性、实时性,且有效解决了计算量问题。

In the paper, we put forward a method to estimate the 6 Dof pose of an object based on RGB image and Inception-ResNet V2 . On the basis of image segmentation, we used the mask to estimate the pose of an object with an real-time output . The traditional algorithm is usually based on RGB-D images, but the RGB-D cameras are always big, expensive and use a lot of power. Ccompared with the RGB cameras, they are not fit for industrial situation and robot, so we used RGB image to estimate the pose. Training a DNN needs a large number of data, for the purpose of training a better Net, we found a way to produce the datasets. Expriments show that, our method has a lower position error and orientation error, the time of estimation is about 0.1s. We have achieved better accuracy and real-time performance, and solved the problem of a larger amount of calculation.