基于卷积神经网络的抓取框检测方法

Grasping Box Detection Method Based on Convolutional Neural Network

服务型机器人抓取的难点在于物体形状不规则、物体位姿随机性以及背景环境复杂。针对此问题,提出一种基于卷积神经网络的机器人抓取方法。该方法以深度图信息作为输入,采用轻量级卷积神经网络将抓取质量、抓取方向和抓取角度映射为热图,根据质量热图中的峰值生成候选抓取框,并从中选取最优抓取框。为验证论文研究方法有效性,基于Cornell抓取数据集进行训练,使用IntelRealSenseD415i深度相机和UR5机械臂搭建实验平台,在真实场景下对随机摆放的物体进行抓取实验。对比试验表明,在Cornell数据集上的准确率和检测速度均有提高,分别达到88.2%和21.0 ms,对数据集之外的物体,抓取成功率达到86%。综上所述,该方法能够快速、精确地对多个物体分别生成抓取框,满足抓取任务的需要。

The difficulty of service robot is that the object shape is irregular,the object pose is random and the background en-vironment is complex.To solve this problem,a robot grasping method based on convolutional neural network is proposed.In this method,the depth map information is used as input,and the grasping quality,grasping direction and grasping angle are mapped in-to a heat map using lightweight convolutional neural network.The candidate grasping boxes are generated according to the peak val-ues in the mass heat map,and the optimal grasping boxes are selected.In order to verify the effectiveness of the research method in this paper,the training is conducted based on the Cornell capture data set,and the IntelRealSenseD415i depth camera and UR5 ma-nipulator are used to build the experimental platform,and the random objects are captured in the real scene.The comparison test shows that the accuracy and detection speed are improved on Cornell data set,reaching 88.2%and 21.0 ms,respectively.For ob-jects outside the data set,the success rate of grasping reaches 86%.To sum up,this method can generate grasping frames for multi-ple objects quickly and accurately,and meet the needs of grasping tasks.