An improved model for target detection and pose estimation of a teleoperation power manipulator

• Published in: Frontiers in neurorobotics 2023-08, Vol.17, p.1193823-1193823
• Main Authors: Xie, Li, Huang, Jiale, Li, Yutian, Guo, Jianwen
• Format: Article
• Language: English
• Subjects: Algorithms; Augmented reality; camera; Cameras; Deep learning; Methods; Motion capture; Neural networks; Neuroscience; Nuclear power plants; pose estimation; Radiation; Robots; target detection; teleoperation power manipulator; Virtual reality
• ISSN: 1662-5218; 1662-5218
• DOI: 10.3389/fnbot.2023.1193823

A hot cell is generally deployed with a teleoperation power manipulator to complete tests, operations, and maintenance. The position and pose of the manipulator are mostly acquired through radiation-resistant video cameras arranged in the hot cell. In this paper, deep learning-based target detection technology is used to establish an experimental platform to test the methods for target detection and pose estimation of teleoperation power manipulators by two cameras. In view of the fact that a complex environment affects the precision of manipulator pose estimation, the dilated-fully convolutional one-stage object detection (dilated-FCOS) teleoperation power manipulator target detection algorithm is proposed based on the scale of the teleoperation power manipulator. Model pruning is used to improve the real-time performance of the dilated-FCOS teleoperation power manipulator target detection model. Compared with the original model, the proposed model was experimentally proved to reach basically the same precision within a shorter inference time (only 58% of that of the original model). To improve the detection speed for the key points of the teleoperation power manipulator, the keypoint detection precision and model inference speed of different lightweight backbone networks were tested based on the SimpleBaseline algorithm. MobileNetv1 was selected as the backbone network to perform channel compression and pose distillation on the upsampling module so as to further optimize the inference speed of the model. The experimental results show that the compressed model basically retains the precision of the original model and that its inference time is 48% of that of the original model.