【摘要】 针对物体姿态随机、外形不规则、抓取环境复杂等机械臂抓取问题，提出一种基于RGB-D的像素级抓取检测方法。首先，对Cornell数据集基于像素级重标注，并生成抓取标签；然后，提出一种多残差提取卷积神经网络(CRE-Net),增强网络特征提取效果；最后，搭建仿真抓取系统，进行算法验证。实验结果表明：位姿检测精度达到93.99%,在对抗性抓取中，单物体抓取成功率为94.0%,多物体抓取成功率为73.3%。更多还原

【Abstract】 Aiming at the grabbing problems of robot arm, such as random object attitude, irregular shape and complex grabbing environment, a pixel-level grabbing detection method based on RGB-D is proposed.Firstly, the Cornell dataset is re-labeled based on pixel-level, and grabbing tags are generated.Then, a convolutional neural network for multiple residual extraction(CRE-Net)is proposed to enhance the effect of network feature extraction.Finally, a simulation grabbing system is built to verify the algorithm.The experimental results show that the precision of pose detection reaches 93.99 %.In adversarial grabbing the success rate of single object grabbing is 94.0 %,and the success rate of multi object grabbing is 73.3 %.更多还原