基于多模态深度神经网络的抓取检测方法

Grasp Detection Method Based on Multi-modal Deep Neural Network

针对机器人抓取检测任务中对未知物体抓取检测精度低的问题,提出了一种多模态深度神经抓取检测模型。首先,在RGB和深度两个通道中引入残差模块以进一步提升网络的特征提取能力。其次,引入多模态特征融合模块进行特征融合。最后,通过全连接层回归融合特征得到最佳抓取检测结果。实验结果表明,在Cornell抓取数据集上,本文方法的图像拆分检测精度达到95.7%,对象拆分检测精度达到94.6%。此外,还通过消融实验证明了引入残差模块可以提高网络抓取检测性能。

A multi-modal deep neural network grasping detection network was proposed to address the issue of low accuracy in robot grasping detection tasks for unknown objects.Firstly,residual modules were introduced in both RGB and depth channels to further enhance the feature extraction capability of the network.Secondly,a multimodal feature fusion module was introduced for feature fusion.Finally,the best grasping detection result was obtained by fusing features through fully connected layer regression.The experimental results demonstrate that the algorithm proposed achieves a precision rate of 95.7% for grasping and 94.6% for object segmentation on the Cornell dataset.In addition,it has been demonstrated through ablation experiments that introducing residual modules can improve the performance of network crawling detection.