Laser & Optoelectronics Progress, Volume. 60, Issue 6, 0615005(2023)
Robot Dynamic Object Positioning and Grasping Method based on Two Stages
Figures & Tables(15)
Fig. 6. Experimental results of object pose estimation. (a) Original images; (b) pose estimation result images
Fig. 8. Robot object grabbing diagrams. (a) Overall pictures; (b) partial pictures
Table 1. First seven-layer convolution structure of improved VGG
View tableTable 1. First seven-layer convolution structure of improved VGG
Network layer Input size Kernel size Output size Conv1+BN+ReLU 3×640×640 3×3 64×640×640 Conv2+BN+ReLU 64×640×640 3×3 64×640×640 MaxPooling 64×640×640 2×2 64×320×320 Conv3+BN+ReLU 64×320×320 3×3 128×320×320 Conv4+BN+ReLU 128×320×320 3×3 128×320×320 MaxPooling 128×320×320 2×2 128×160×160 Conv5+BN+ReLU 128×160×160 3×3 128×160×160 Conv6+BN+ReLU 128×160×160 3×3 128×160×160 Conv7+BN+ReLU 128×160×160 3×3 128×160×160 MaxPooling 128×160×160 2×2 128×80×80
Table 2. Comparison of model parameters, model prediction time, and segmentation accuracy
View tableTable 2. Comparison of model parameters, model prediction time, and segmentation accuracy
Method Size /MB Average running time of test image /s mIoU /% UNet［ 18 ］94.97 2.23 95.6 PSPNet［ 19 ］9.31 0.36 83.5 BiSeNet［ 20 ］52.52 0.68 91.5 ICNet［ 22 ］31.52 0.96 95.3 Fast-SCNN［ 23 ］4.60 0.32 88.6 BiSeNetV2［ 21 ］20.47 0.51 95.1 Proposed method 19.08 0.41 96.8
Table 3. Accuracy and prediction speed comparison of multi-scale perception layers
View tableTable 3. Accuracy and prediction speed comparison of multi-scale perception layers
Method mIoU /% Average running time of test image /s Proposed method（one MSPL） 91.5 0.32 Proposed method（two MSPL） 95.1 0.36 Proposed method（three MSPL） 96.8 0.41 Proposed method（four MSPL） 97.2 0.48
Table 4. Pose data table of item grabbing point
View tableTable 4. Pose data table of item grabbing point
Thing Posture Predict time /s Coordinate Angle /（°） Gel pen （545，433） 120.23 0.015 （526，517） 58.34 Paper knife （422，485） 124.55 0.014 （463，491） -29.07 Remote control （368，432） 128.56 0.015 （538，331） 31.20 Scissors （407，420） 122.96 0.016 （544，399） -34.17 Screw （544，379） 127.37 0.014 （654，522） 59.88 Nut （695，217） 82.24 0.016 （537，318） 84.25
Table 5. First group of grasping test results
View tableTable 5. First group of grasping test results
Thing Number of experiments Number of successful crawls Success rate /% Remote control 30 29 96.7 Paper knife 30 27 90 Scissors 30 28 93.3 Screw 30 24 80 Nut 30 26 86.7 Gel pen 30 28 93.3
Table 6. Second group of grasping test results
View tableTable 6. Second group of grasping test results
Thing Number of experiments Number of successful crawls Success rate /% Remote control 30 29 96.7 Paper knife 30 28 93.3 Scissors 30 28 93.3 Screw 30 26 86.7 Nut 30 26 86.7 Gel pen 30 28 93.3
Table 7. Third group of grasping test results
View tableTable 7. Third group of grasping test results
Thing Number of experiments Number of successful crawls Success rate /% Remote control 30 30 100 Paper knife 30 28 93.3 Scissors 30 28 93.3 Screw 30 28 93.3 Nut 30 28 96.7 Gel pen 30 29 96.7
Tools
Get Citation
Copy Citation Text
Yuebo Meng, Qi Huang, Jiuqiang Han, Shengjun Xu, Zhou Wang. Robot Dynamic Object Positioning and Grasping Method based on Two Stages[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0615005
Paper Information
Category: Machine Vision
Received: Dec. 27, 2021
Accepted: Jan. 27, 2022
Published Online: Mar. 16, 2023
The Author Email: Meng Yuebo (mengyuebo@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20