Research On Manipulator Perception Technology Based On Active Stereo Vision

In recent years,with the rapid development of artificial intelligence technology,the combination of robot and visual technology has become a research hotspot in the academic field.Vision technology can improve the autonomy and accuracy of robot work,expand the workspace and scene of robots,and is an important direction for future development.Active stereoscopic cameras are a combination of active projection and binocular cameras,which have the advantages of strong anti-interference ability,high accuracy,and strong real-time performance.Active stereo vision technology includes stereo information acquisition technology and image information interpretation technology,which can provide target category information,distance information,position information,and posture information for robotic arm systems working in dynamic scenes.Active stereovision technology can not only be applied in industrial environment to support industrial robot arms to complete various production tasks.It can also be applied to the space environment to assist the space manipulator to complete important tasks such as in-orbit formation and space debris cleaning.Therefore,the development and research of robotic arm sensing technology based on active stereo vision has important research significance and application value.In this dissertation,aiming at the information needed in the perception process of robotic arm systems based on active stereo vision and the problems existing in key algorithms,the relevant visual algorithms have been studied according to task clues and logical order,and advanced results have been achieved.The research results have theoretical significance and application value.(1)In order to generate dense depth maps to obtain distance information,an active stereo matching algorithm based on hierarchical recurrent neural networks was proposed to address the error matching problem of stereo matching algorithms under the influence of non-textured regions,noise,and uneven illumination.Using binocular images as input,a disparity map(depth map)is generated through four stages: feature extraction,local feature attention,correlation cost product,and disparity hierarchical cycle.The designed feature extractor enhances dim feature learning through multi resolution residual linking,improving the robustness of the algorithm to light changes.This method combines the local attention mechanism with local window feature maps to reduce the impact of uneven noise intensity distribution and enhance the positional correlation of features.Finally,an iterative contrast reconstruction loss is proposed to overcome the error dependence of noise and distance.Experiments on public datasets and active stereoscopic images show that the algorithm has advanced matching effects and can generate accurate and detailed disparity maps.(2)In order to accurately segment objects in an image to obtain location information,an image segmentation algorithm based on multiscale extended convolutional neural networks was proposed to address the error segmentation problems caused by grid artifacts,uneven illumination,and similar structures in coder decoder semantic segmentation algorithms.Using a single frame RGB image as input,a semantic segmentation architecture of "encoder attention decoder" is adopted to generate a object segmentation mask image.This method uses an extended convolutional residual network as an encoder to reduce the impact of mesh artifacts and improve the segmentation accuracy of noisy images.Secondly,a channel attention mechanism is introduced to recalibrate the feature response through the squeeze and stimulation modules,while focusing on the two branches of the network,namely,deep and shallow features,to improve learning effectiveness.Finally,a parallel hole pyramid pooling structure is designed to further enhance context information and improve the effect of multiscale object segmentation.Experiments on public and self-made datasets show that the algorithm outperforms similar algorithms and can segment more complete and clear object masks.(3)In order to accurately obtain target pose information,a rigid body 6D pose estimation algorithm based on point cloud fractal fusion neural network was proposed to improve the accuracy and computational speed of pose estimation algorithms,and make better use of color and geometric features.Using a single frame RGB D image as input,using a heterogeneous information dense fusion architecture,the target pose estimation parameters are output through four stages: color feature extractor,geometric feature extractor,pixel level dense fusion network,and pose refinement.This method uses multi scale extended convolutional neural networks to embed color features for dense fusion structures,adapting to multi scale targets and complex scenes.The designed point cloud feature extractor effectively extracts local and global features,enhancing the learning of local geometric information.Finally,a fractal pyramid structure of point clouds is designed to densely fuse multi resolution point clouds with different densities and color features,improving the refinement effect of pose.Experiments on publicly available pose estimation datasets show that the algorithm has high accuracy and can quickly estimate the target pose during reasoning.After connecting the algorithms in this dissertation in series,the validity and feasibility of the vision algorithm at each stage are verified through robotic arm grasping experiments.In this dissertation,aiming at the information needed in the perception process of robotic arm systems based on active stereo vision and the problems existing in key algorithms,the relevant visual algorithms have been studied according to task clues and logical order,and advanced results have been achieved.The research results not only have enlightening significance at the theoretical level,but also have certain practical application prospects.