Research On The Recognition Of 3D Multi-object Images Based On Robot Arm Platform

With the advent of artificial intelligence,traditional industrial robots have gradually developed into intelligent robots which had corresponding scene perception functions.The combination of industrial robots with image processing technology has become an irresistible development trend.In industrial production line of some scenario because of low efficiency of manual sorting parts,using computer vision technology to identify parts technology has become a very valuable research direction.The visual robot used the corresponding visual system to identify the corresponding artifacts,then drive mechanical arm to grab,which greatly improve the production efficiency.Therefore,in the process of assembly,sorting,welding and other industrial production,the study of parts image recognition becomes very meaningful.The recognition of a three-dimensional(3D)multi-object images system based on robot and image processing technology is proposed.First,the hardware was selected according to the actual working scene of the system;then the independent joint control of the robot arm was analyzed to perform the recognition.On the basis of rigid body dynamics theory,a nested PD controller with feedforward compensation was designed using MATLAB Simulink module for simulation.By contrast,the tracking error before and after feedforward compensation were greatly reduced which laid a foundation for the development of the further robot to grasp accurately.The binocular vision system was designed for calibration by using MATLAB toolbox to obtain the internal and external parameters and structural parameters of the left and right cameras.Bouguet algorithm was used for stereo correction of the visual system,and the co-alignment of left and right images was achieved after correction.Because the traditional part detection technology need to design manual features which the algorithm robustness was poor,the deep convolutional neural network was used to identify the parts.The whole algorithm is improved based on SSD framework,adding feature fusion module to combine deep features and shallow features to strengthen the detection of small targets.The algorithm also adds k-means algorithm to self-learning candidate box.The experimental results show that the average detection accuracy of the improved SSD algorithm reaches at 97.11%,which is 8.35% higher than the original SSD algorithm and the detection speed reaches at 26.46 fps to achieve the effects of the real-time.For occluded parts,the exclusion loss technology is used to solve the occlusion problem,then the occlusion subset is set according to the occlusion rate.The experiment shows that the m AP increased by 4.59% after adding the repulsion,which greatly enhanced the robustness of detecting the occluding parts.Finally,a stereo matching algorithm based on SSD was proposed to obtain the depth of the parts,and the Io U(Intersection over Union)index was used to make rough matching toward the prediction boxes of left and right images;whereas,the Census stereo transform was used for precise matching.The positioning experiment was set up;the error of the X,Y,Z direction were found [-0.11,0.13],[-0.21,0.21],and [-0.16,0.22],respectively,which meet the requirements of grasp robot accurately.