Improved Visual SLAM Algorithm in Factory Environment

Aiming at the high positioning accuracy of automatic guided vehicle (AGV) in industrial scene, an improved visual simultaneous localization and mapping (VSLAM) algorithm is proposed. In the front end of the algorithm, the stereo images are captured using a binocular camera and sub-pixel-level matching point pairs are obtained through binocular matching algorithm. And the 3D information of these feature points is calculated in the camera coordinate system. Then the pose transformation is acquired using RansacPnP algorithm based on the 3D-2D matching point pairs. And the transformation is used as the initial value to do the local optimization by minimizing re-projection errors. Additionally, an uncertainty model is proposed based on the Hamming distance between the descriptors of matching point pairs. The model provides the information matrix for the constraints in the local optimization, which can improve the positioning accuracy. In the back end of the algorithm, reliable closed-loop information of artificial landmarks is detected using the monocular camera shooting vertically upwards, and then the global pose optimization is conducted. According to the motion model and working scene of the AGV, an optimization method based on global plane constraint is proposed to reduce the error of SLAM system. In the experiment, the accuracy of binocular odometer in the front end of the proposed algorithm is compared with the odometers in ORB-SLAM2 and libviso2 algorithm using the offline KITTI dataset. The online experiments in the factory environment are conducted to test the real localization accuracy and robustness of the whole system. The results show that the algorithm is of good overall performance, and its positioning errors are within 10cm, while the positioning frequency can reach 20Hz, which can meet the industrial requirements.