Study On Visual Pose Measurement And Servoing Control Of Palletizing Robot Based On Neural Network

Based on timber palletizing with robot in the wood processing industry, robot vision technology is applied to gives the robot visual sense and makes the robot have rapid response to the change of surrounding environment and operation objects and guide the robot to carry and pallet. So the automation and intelligence of robot are improved and the industrial production efficiency is guaranteed. The image acquisition, image filtering and feature extraction of robot vision system were studied. The 3D pose measurement of object with the image features was studied. The robot visual servoing control technology was studied with the pose as the control input to control robot to carrying and palletizing. The main conclusions are as follows:(1)Probabilistic Neural Network has excellent classification ability. Based on the pixel gray value of noisy image, Rank-ordered mean of the absolute deviations (ROMEAD) and Rank-ordered Median of the Absolute deviations (ROMDAD) were established. ROMEAD vector was used to calculate the location of more than 70% noise. ROMDAD vector was used to calculate the closeness degree of current pixel value and the neighbour for distinguishing between salt & pepper noise pixels and non-noise pixels effectively. ROMEAD and ROMDAD were as the PNN inputs. With the training, the PNN could detect and isolate the Gaussian noise and Salt & Pepper noise effectively which was filtered by mean filtering and median filtering respectively. Radial basis function neural network has excellent learning ability. The mean filtering image and median filtering image were as input and fused by RBFNN. More useful image information and edges were reserved.(2)In the traditional line feature extraction methods, canny operator and Hough transform had good extraction effect. But canny operator had poor anti-noise capacity and subjected to threshold value. So the adaptability of canny operator was poor. Hough transform was affected by noise small, but had better robustness, so was widely used. Hough transform was chosen to extract the line feature in this study.(3)The vanishing point is an intersection point of a group of parallel lines in the perspective projection. According to the properties of vanishing point, it contains intrinsic parameters of camera. On the basis of single 2D image, focal length and optical center of camera were calculated. The rotation matrix R in the relation matrix between camera and scene was calculated with the vanishing point which was formed by the pairwise orthogonal and non-coplanar three groups of lines in the pair of matching images. And then based on similar triangle principle, the translation component t was calculated by the same line in the matching relation of coordination system between the coordination system based on optical point and the coordination system based on vanishing point.(4)Based on dual-quaternion, the rotation matrix R and position vector t in the transformation relation between camera coordination system and robot end effector coordination system was calculated simultaneously. This method could eliminate transmission errors during calculation. Error function was built with the measured values and theoretical values of position quaternions and direction quaternions on the object. And then Lagrange function was built by Lagrange multiplier with the constraint conditions. Based on Hopfield neural network, the optimal solution of error function was solved.(5)The 3D pose of object as the input of robot system, based on the position-based visual servoing, the backstepping controller was designed by backstepping. Control law and adaptation law were chosen to decrease the effect of errors and others uncertainties on robot control and ensure the robot system global asymptotic stability. Based on RBFNN, robot was controlled without accurate robot model. Robust control law was designed to improve the robot resistivity against error and interference and system robustness.