Research of Workpiece Pose Estimation with DeepLabV3 Neural Network
For industrial parts with few surface features,feature-based matching algorithms cannot be used,and traditional tem-plate-based matching algorithms are not robust in scenes which are varying lighting and have chaos background.So a method based on DeepLabV3 network that combine with traditional algorithms is proposed,which greatly improves the matching results and detection accuracy in complex background.First,with the Humoment and the minimum bounding rectangle method we deter-mine the center position and rotation angle under the DeepLabV3 network.In the second step,we use the adaptive Harris corner detection combined with the five-point method to complete the rapid calibration of the hand-eye camera,and finally complete ex-periment under the AUBO robotic arm.In the positioning experiment,the positioning error is within 0.5mm,and the results show that the algorithm has better performance in complex scenes and non-uniform illumination.
NETL
NSTL
万方数据
