Visual tracking control of SCARA robot system based on deep learning and Kalman prediction method

Abstract

When the workpiece moves, the existing industrial machine vision tracking system has some problems, such as low accuracy, poor recognition effect and so on. In this paper, optimising convolution neural network based on stochastic gradient descent is used for motion foreground segmentation and tracking object, to realise high-speed and high precision of recognition of the moving object. The hard triggered capture system ensures the time interval of the image sequence. And an improved image sequence speed detection algorithm is proposed, combined with the Kalman prediction equation to predict the coordinate position of the workpiece in the nonlinear motion state in order to compensate the time delay error of the visual servo system. Finally, the experiment is carried out on visual tracking robot system with the belt conveyor. The results show that the maximum tracking speed can reach 90 mm/s, and the tracking accuracy error is less than 1 mm. Although facing sudden speed changes in conveying blet, the robot system can be stable in several milliseconds which prove that this method has the characteristics of high reliability and good tracking accuracy.