基于改进CenterNet的机械臂抓取检测

Robotic grasping detection based on improved CenterNet

在机械臂的抓取检测中,基于Anchor-based的方法需要考虑很多超参数的选择和设计,难以有效提高算法整体性能。针对该问题,本文将抓取检测转换为关键点检测问题,基于CenterNet提出一种改进的抓取检测模型。首先,该模型重点解决寻找抓取框中心点的问题,其本质是对中心点进行关键点估计,从而降低抓取检测的复杂程度;其次,采用HourglassNet神经网络提取深层特征;然后,为了使模型能聚焦抓取检测中物体的重要特征,设计一种视觉注意力机制;最后,生成关键点的高斯热力图和嵌入式向量,并将抓取框的中心点设置为检测出的关键点位置,将抓取框的中心点偏移量,长,宽以及旋转角分别设置为嵌入式向量中的值,从而有效解决了这些超参数的设置问题。研究结果表明:在康奈尔抓取数据集上进行图像分割和对象分割实验,准确率分别达到了98.3%和96.7%;本文方法通过计算获得一些较优的参数而不是采用先验参数,与其他基于Anchor-free的方法相比,提高了检测精度,而且其计算精度超过一些常用的基于Anchor based的模型的精度。

In the robotic grasping detection,anchor-based method needs to consider the selection and design of many super-parameters,which is difficult to effectively improve overall performance of algorithm.In order to solve this problem,the grasp detection was transformed to the key point detection problem,and a grasp detection model based on improved CenterNet was proposed.Firstly,the problem of finding center point of grab box was focused on of the model and the essence was to estimate the key points of center point,so as to reduce the complexity of grab detection.Secondly,to extract deep features,deep convolutional neural network HourglassNet was used.Thirdly,a visual attention mechanism was proposed so that the model pay more attention to the important features of object in detection,which was useful to key point detection and regression calculation.Finally,Gaussian heat map and embedded vector of the key points were generated and the center point of the grab box was set as the detected key points.The center point offset,length,width and rotation Angle of grab box were set as the values in the embedded vector to effectively solve the problem of setting these super-parameters.The results show that the method in this paper obtains better parameters through calculation rather than using prior parameters.The accuracy of image segmentation and object segmentation on Cornell Grasp Detection Dataset is 98.3%and 96.7%,respectively.Compared with other anchor-based methods,it improves the detection accuracy and can exceed some commonly used Anchor based models.