LSFP-Tracker: An Autonomous Laser Stripe Feature Point Extraction Algorithm Based on Siamese Network for Robotic Welding Seam Tracking

Abstract

Laser vision based welding seam tracking has become the key technology to realize intelligent robotic welding, and extracting the feature point of welding seam is an essential step. Aiming at the low flexibility of classical methods, this paper proposed a laser stripe feature point tracker (LSFP-tracker) based on siamese network for robotic welding seam tracking. Firstly, a lightweight feature extraction network (LSFENet) is constructed to produce high resolution feature maps that contain both low-level geometric details and high-level semantic information, which can ensure model's tracking accuracy and anti-noise capability. Moreover, a feature refinement module is designed, which can enhance the tracking stability on specific seams through channel-wise emphasizing useful features. Subsequently, stripe's curvature and three preset templates were introduced to adaptively detect the initial feature point and corresponding bounding box, which is helpful to enlarge the applicability scope of LSFP-tracker. The experimental results show that the LSFP-tracker can maintain satisfying tracking performance on various welding seams, such as butt, V-groove, multi-pass seams, etc., even under the interference of welding noise. And the comparison tests prove the advantage of LSFP-tracker, which has great application potential in practical welding.