Binocular Positioning Pipeline Grabbing System Based On Deep Learning

In order to improve industrial production efficiency,researchers are introducing intelligent algorithms and computer vision technology into industrial equipment to overcome the shortcomings of traditional control methods.This thesis studies the task of the robotic arm sorting objects on the assembly line,uses binocular vision and deep learning object detection technology to detect and locate moving objects,and controls the robotic arm through an embedded terminal to grab the target.In order to conduct follow-up experiments,this thesis designs an industrial simulation platform and completes relevant preparatory work,including building an industrial assembly line platform,selecting various hardware,determining component connection relationships,and calibration and correction tasks for binocular cameras.Through modeling and relational reasoning,the relationship between spatial coordinates and image pixel coordinates is determined,and the Zhang’s calibration method is selected to complete the calibration and correction tasks,and to ensure that the obtained binocular views are normal,undistorted and horizontally corresponding.In order to realize real-time detection of dynamic pipeline objects,this thesis proposes the YOLOX-GRC depth detection algorithm to complete the detection of target objects on binocular images.This thesis first builds a dataset for deep learning based on the built system.Then,based on the YOLOX model,the characteristics of the industrial assembly line recognition scene are improved.Firstly,in view of the limited computing power of edge devices,branch reduction is performed on the YOLOX model,and the detection branch with less influence is removed.Then,aiming at the situation that it is difficult to recognize small targets in the detection environment,the FOCUS module in the original model is modified to the GHOST module to reduce information loss.Aiming at the characteristics of the large background in the scene,the RFB module and the attention mechanism module are introduced into the model to enhance the receptive field of the model features and highlight the key detection areas of the model.Finally,the number of each feature extraction module in the backbone network is adjusted to ensure optimal detection efficiency.Finally,the YOLOX-GRC model suitable for the research environment of this thesis is proposed.On the constructed industrial simulation data set,the accuracy reaches 81.6%,and the inference time is only 19.6ms,which far exceeds the original model and the commonly used neural network detection model..In order to realize the spatial positioning of pipeline objects,this thesis proposes the YOLOX-3d algorithm based on the YOLOX-GRC target detection model,which combines binocular vision positioning logic and constraints between binocular corresponding points for 3D point recognition of targets.Algorithms are deployed on embedded platforms and optimized using Tensor RT technology.The experimental results show that YOLOX-3d only takes 42.63 ms to complete a calculation on the embedded platform.Among the 20 simulated grabbing tasks,18 were successfully completed,and the other two were processed by the abnormal detection and exclusion module to avoid false detection.The resulting system exception.