基于多模态深度神经网络的抓取检测方法

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,本文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。接着,引入多模态特征融合模块进行特征融合。最终通过全连接层回归融合特征以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,本文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。

Grasp Detection Method Based On Multi-modal Deep Neural Network

This paper proposes a multi-modal deep neural network grasping detection network, which is mainly aimed at the low accuracy of grasping and detecting unknown objects in the task of manipulator grasping detection. Firstly, the residual module is introduced into the RGB and depth channels, and the performance of the network is enhanced. Then, a multimodal feature fusion module is used to merge the feature. Finally, the fusion feature is returned to the full connection layer to get the best capture test results. The experiment results show that the precision of the proposed algorithm is 95.7% and the precision of object segmentation is 94.6% on Cornell dataset.In addition, this paper proves that the introduction of the residual module can improve the quality of network capture through ablation experiments.