Error analysis of dynamic localization tests based on binocular stereo vision on litchi harvesting manipulator

Abstract: The random error caused by disturbance is the bottleneck of vision positioning of harvesting manipulator. In order to improve the work-efficiency and positioning accuracy of litchi harvesting manipulator, the precise positioning of litchi picking point was studied. Firstly, a binocular stereo vision and an added vibration table to simulate external force disturbance were used to detect the picking point coordinates. The experiment was conducted using the mechanism and the vision-positioning experimental platform based on binocular vision in laboratory. The cameras used in the test and that on the manipulator were the same. The vibration table created perturbation in three coordinate directions. The corresponding perturbation was close to that when conducting outdoor work in a dynamic environment (such as mechanism vibration or wind). A calibration plate was used as a reference to detected positions in the horizontal direction, and a laser rangefinder was used to measure ground truth in the depth direction. The accuracy reported by the specs of calibration plate and laser rangefinder were ±0.2 mm and ±1 mm, respectively. After measuring the ground-truth position, the errors were calculated. Secondly, an error analysis scheme was proposed. The errors resulting from vision positioning system were divided into original positioning and random positioning errors according to the change rules. The positioning errors under static conditions were regarded as original positioning errors. Random positioning errors were resulted from the influence of unknown external forces. Finally, the quantitative analysis and evaluation of the errors were separately determined by statistical methods. The results showed that the variation trend of original positioning errors was basically consistent with random positioning errors in the distance of 600 to 1000 m. The maximum errors obtained were 58.8 in vision depth direction and 17.3 mm in horizontal direction. The positioning precision in horizontal direction was high. Errors in vision depth direction were considerably larger than errors in horizontal direction. The original error achieved a narrow confidence interval. The original positioning errors in vision depth direction and distances were linear, while the errors in horizontal direction were nonlinear. The random positioning errors under dynamic environment were normally distributed. The random positioning errors in vision depth direction had no obvious connection with errors in horizontal direction. The random positioning errors in vision depth direction, which were greatly influenced by the disturbance, showed a low positioning accuracy and were worse than the errors in horizontal direction. So the disturbance restraint in vision depth direction should be improved. The experiment results can be used to calibrate vision system and layout the vision positioning scheme under dynamic environment. Meanwhile, a mathematical model of error tolerance was established. As the existing end-effectors cannot conduct error tolerance for random error, it also provided a theoretical basis for error -tolerant design of fruit-picking end-effector. Indoor positioning and grasping experiments were conducted for litchi using the manipulator based on the binocular vision. The picking success rate was over 90%. The results verified the applicability of the error-tolerant design. In summary, a precise positioning measurement was proposed based on institutions and vision positioning, and random positioning error could be compensated for by the mechanism of error tolerance.