Autonomous weld seam tracking under strong noise based on feature-supervised tracker-driven generative adversarial network

Strong noise from complex welding condition such as arc light and splashes lead to high tracking error in vision-based seam tracking. To solve this problem, this paper proposes an autonomous seam tracking method based on DCFnet. A feature-supervised tracker-driven generative adversarial network (FT-GAN) is introduced to repair the noise interfered laser stripe images. A feature supervision module and feature selection module are designed in the feature extraction process of the encoder. In addition, the DCF tracking response loss is added to the loss function, guiding the tracking-oriented feature restoration. During tracking process, images are first repaired images by FT-GAN and automatically tracked with DCFnet for laser feature point. In order to promote tracking performance and reduce extract calculation, the model parameter updating and image inpainting frequency is controlled by the response peak side lobe ratio (PSLR). In experiments, the tracking speed reaches up to 15 fps, and the average error is controlled within 0.236 mm. Test results prove that this seam tracking method performs well in accuracy, efficiency and robustness over traditional methods.