基于Faster-RCNN的智能机器人自动抓取系统研究  

Research on the Automatic Grasping System of Intelligent Robot Based on Faster-RCNN

作　　者：张浩 Zhang Hao(Gansu Forestry Polytechnic,Tianshui 741020,China)

机构地区：[1]甘肃林业职业技术学院,天水741020

出　　处：《办公自动化》2024年第11期21-25,共5页Office Informatization

摘　　要：为实现复杂场景、多模态目标物体的自适应定位、抓取,文章设计一种基于改进Faster-RCNN算法的智能机器人自动抓取系统。首先,以DSFPN算法改进Faster-RCNN的特征提取方法,通过低、高层次语义信息融合,改进小尺寸目标特征提取精度;并采用CDN通过可变形卷积的逐层叠加,提升复杂环境下多模态目标体检测精度,而后,引人NP层对训练进行自适应反馈调节,以控制候选区数量,优化目标检测的效率,由此,基于ROS机器人操作系统,利用Kinect相机获取目标物体的位姿,通过手眼标定数据完成目标物体位姿转换,实现自适应抓取。最后,通过对比实验,验证基于改进Faster-RCNN算法的抓取系统,目标检准率、检全率、综合检测率分别达到93.68、92.71%、93.06%,尤其对小尺寸目标的检测准确性更优,可满足多变环境下不同目标物体的抓取需求。In order to realize the adaptive positioning and grasping of multi-modal objects in complex scenes,an intelligent robot automatic grasping system based on improved Faster-RCNN algorithm is designed.Firstly,the Faster-RCNN feature extraction method is modified by DSFPN algorithm,and the feature extraction accuracy of small-scaled targets is improved by the low-level and high-level semantic information integration;and the CDN is used to improve the detection accuracy of multi-modal targets in complex environment through layer by layer superposition of deformable convolution,and then,the NP layer is introduced to adjust the training adaptively,so as to control the number of candidate regions and optimize targets.Therefore,based on the ROS robot operating system,Kinect camera is used to obtain the pose of the target object,and the hand eye calibration data is used to complete the pose conversion of the target object,so as to achieve the adaptive grasping.Finally,through the comparative experiments,it is verified that the target detection rate,full detection rate and comprehensive detection rate of the grasping system based on the improved Faster-RCNN algorithm are 93.68%,92.71%and 93.06%respectively,especially for the detection accuracy of small scaled targets,which can meet the grasping requirements of different objects in the changeable environment.

关 键 词：复杂场景 Faster-RCNN算法 多模态目标 自适应抓取 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置] TP391.41[自动化与计算机技术—控制科学与工程] TP273