Robotic Seam Tracking System Combining Lightweight Segmentation Network Design and ADMM-Based Structured Pruning

Abstract

To achieve precise real-time tracking of weld seams, designing a weld seam tracking system that employs the laser vision sensor is essential. Throughout the welding process, the occurrence of welding noise is inevitable, which subsequently obscures critical weld feature point and diminishes welding accuracy. Consequently, we have proposed an image segmentation method employing a lightweight segmentation network to separate this noise. We first establish its structural design by incorporating the UNet as the initial network and introducing a novel V-layer structure. Then we employ Alternating Direction Method of Multipliers (ADMM) optimization principles to prune and accelerate the initial network. Furthermore, we have designed a seam tracking system to validate our proposed method. The experiment results, when integrated into the Efficient Convolution Operators for tracking(ECO) algorithm, demonstrate an average localization error for lap-type workpieces is 0.19 mm, for butt-type workpieces it is 0.053 mm, and the maximum localization error for both types remains within 0.35 mm, fulfilling the precision and real-time performance criteria set by the weld seam tracking system. Ultimately, our proposed lightweight segmentation method is not limited to the welding realm, serving as a valuable reference for segmentation applications in various other industrial fields.