Detection And Positioning Of Grab Target Based On Deep Learning

In recent years,with the rapid development of national economy and science and technology,the combination of social production and life with robot technology has become more and more extensive;at the same time,as computer vision processing algorithms and computing hardware based on deep learning develop rapidly,it is robots and intelligent robots integrated with deep learning that have become a research hotspot and focus.Aiming at the problems of low detection accuracy of the grasping angle of the target and inaccurate positioning of the grasping point by the robot,this study proposes a two-stage detection method for processing robot visual information based on a convolutional neural network to achieve quick grabbing posture detection and precise positioning of grabbing points.The first stage of detection uses a multi-grabbing target detection algorithm to complete the multi-grabbing target detection of the grabbing targets in the input picture of the robot system whose purpose is to find the targets to be grabbed among multiple targets.The network outputs each position and category of the target,and outputs the regional feature map of the target to be captured to the next stage.Cut and expand the Cornell data set,dividing the data set into 20 categories according to daily categories,and label the data set according to the Pascal VOC data labeling format.The end-to-end multi-grabbing target detection model is trained.The average accuracy of the algorithm model for grabbing target recognition reaches 83.2%,and the processing time of each image is less than 0.04 seconds.The multi-grabbing target images input by the multi-crawling target detection model output their positions and categories,narrowing the detection range for the next stage of posture detection and positioning,and reducing the impacts of complex environmental backgrounds.In the second stage,the output of the multi-grabbing targets in the first stage is used as an input to perform posture detection and grabbing point positioning on the grabbing targets.Compared with the previous single posture angle and positioning classification or regression model,this research method proposes a regional posture detection and positioning algorithm model,which outputs the posture angle of the grabbing target in the network by classification and uses the regression method output for the grabbing position coordinates.Re-mark the grabbing angle label and grabbing points on the Cornell data set,and train the end-to-end posture detection and positioning detection model,and then input the regional grabbing target output in the first stage into the trained model which can make posture detection and predict the location of the grabbing point for the regional grabbing target images.The model achieved the accuracy of 96.18% and 96.32% on the instance detection and object detection test sets respectively,and the processing time for each image was less than 0.1 second.The model can perform high-accuracy posture angle detection and grabbing point positioning on multiple grabbing targets,and further optimize the real-time detection.In summary,the two-stage detection method for posture detection and positioning of multiple grabbing targets proposed in this study can quickly and accurately detect and locate single or multiple grabbing targets in complex background images in real time with high robustness and stability.