Design And Implementation Of Robotic Arm Pose Estimation System Based On Deep Learning

In industrial manufacturing,accurate and flexible robotic arms have been widely used in various applications,where robotic arms are capable of performing tasks such as assembly operations,object grasping,human-robot interaction,and collision detection.In recent years,vision-based robotic arm control has received increasing attention,as vision-based systems are more informative,flexible,and adaptable to complex and diverse tasks than conventional systems using specific sensors.Vision-based robotic arm control systems are capable of controlling robotic arms based on visual information and can further enable robotic arm grasping.If we want to control a robotic arm,we must first know the position of each joint of the arm,i.e.,the pose of the arm,so the pose estimation of the arm is a necessary step for the control of the arm and is crucial for the implementation of the robotic arm control system.Based on the above background and practical applications,a robotic arm pose estimation system based on deep learning is designed and implemented in the paper,aiming to make the production line more intelligent.The work in this paper can be summarized as follows.(1)The various components of convolutional neural networks and their classical architectures are studied,and the network optimization methods,parameter initialization and network regularization methods of convolutional neural networks are sorted out.The research related to deep neural network compression and acceleration is introduced,and the three aspects of pruning,knowledge distillation,and compact network design are sorted out and summarized respectively,and finally the compact network design approach is chosen to compress the robotic arm pose estimation model.(2)To address the problem of excessive parameters and computational cost of existing robotic arm pose estimation methods,based on Stacked Hourglass Network(SHN)model,we introduce a lightweight Ghost module and Heterogeneous Kernel-Based Convolution(Het Conv)module,and propose deep learning-based robotic arm pose estimation methods: Ghost-SHN and Het Conv-SHN.The methods construct an SHN with two stacks.Ghost-SHN uses Ghost module to replace the standard convolution of SHN,and Het Conv-SHN uses Het Conv module to replace the standard convolution of SHN.Experiments conduct on four typical datasets show that Ghost-SHN and Het Conv-SHN not only reduce the parameters and calculation compared with existing robotic arm pose estimation models,but also have great generalization performance,achieving better accuracy than SHN,and can be better deployed on robotic arms.(3)The overall construction of the deep learning-based robotic arm pose estimation system is implemented,and the viability and application value of the vision system are proved through relevant experimental analysis and practical applications.