中文版 | English
CSpace
FRCNN-Based DL Model for Multiview Object Recognition and Pose Estimation
Zhao, Yongting1; Zheng, Bin1; Li, Haochen1,2
2018
摘要5 types of object pose detect neural networknetworks based on FRCNN(Fast Region-based Convolutional Network) and multi-task framework are presented to improve the success rate of robot grasping and to overcome the problems of low speed, poor applicability and difficulty of samples acquisition caused by a 3D model matching method. The network based on Fast-rcnn add an output layer for the pose estimation and simplify it to a classification problem so that the system can be used to estimate the pose, type and bounding box of object at the same level. The effectiveness of the new presented models are demonstrated in the test experiments under the detection precision of 10, 30 and 45 degrees on industrial PCB and EPFL test samples. Meanwhile, the performance comparisons on the proposed models before are implemented. The range of rotate angle for PCB and azimuth angle for cars in EPFL dataset can be obtained through the computation of model while the accuracy of recognition remains at around 98% and the network can achieve the MPPE(Mean Precision in Pose Estimation) of 97.5%/90.6% 93.6%/88.2% and 89.7%/82.6% under the 3 detect precisions respectively. The experimental results indicate that the feasibility of model to estimate the pose of objects in space of 2 or 3 dimension. That means the model can be applied to the tasks of pose detection on the planar workpiece within the field of industrial handling. © 2018 Technical Committee on Control Theory, Chinese Association of Automation.
语种英语
DOI10.23919/ChiCC.2018.8483556
会议(录)名称37th Chinese Control Conference, CCC 2018
页码9487-9494
收录类别EI
会议地点Wuhan, China
会议日期July 25, 2018 - July 27, 2018