6D姿态实时估计的轻量型YOLO网络

Lightweight YOLO Network for 6D Attitude Estimation in Real Time

针对同类基于卷积神经网络的算法对目标姿态估计存在实时性不足的问题,提出了一种轻量型的6D目标姿态估计算法。该算法通过卷积神经网络对输入的RGB图像进行特征提取,并将目标对象的位置信息映射到三维空间,最后通过目标图像及其三维模型之间的多组2D-3D对应关系,利用PnP和RANSAC方法计算目标的6自由度姿态。实验结果表明,改进后的网络与原网络具有同样优异的检测效果,且检测速度得到了明显的提升,在NVIDIA GeForce RTX 2060 GPU上的运行速度为35 FPS,适用于对处理速度有要求的场合。

Aiming at the problem that similar algorithms based on convolutional neural networks have insufficient real-time performance for target pose estimation, this paper proposes a lightweight 6 D target pose estimation algorithm.The algorithm uses convolutional neural network to extract features of the input RGB image, and map the position information of the target object to the three-dimensional space, and finally through multiple sets of 2 D-3 D correspondence between the target image and its three-dimensional model, using PnP and The RANSAC method calculates the 6-degree-of-freedom attitude of the target.The experimental results show that the improved network has the same excellent detection effect as the original network, and the detection speed has been significantly improved.The running speed on the NVIDIA GeForce RTX 2060 GPU is 35 FPS,which is suitable for occasions that require processing speed.