Robust seam tracking via a deep learning framework combining tracking and detection.

To address the problem of low welding precision caused by possible disturbances, e.g., strong arc lights, welding splashes, and thermally induced deformations, in complex unstructured welding environments, a method based on a deep learning framework that combines visual tracking and object detection is proposed. First, a welding image patch is directly fed into a convolutional long short-term memory network, which preserves the target's spatial structure and is efficient in terms of memory use, with the aim of avoiding some disturbances. Second, we take advantage of features from various convolutional neural network layers and determine weld feature points through similarity matching among multiple feature layers. However, feeding in noisy images causes the tracker to accumulate interference information, which results in model drift. Thus, using a welding seam detection network, the object filter is periodically reinitialized to improve tracking accuracy and robustness. Experimental results show that the welding torch runs smoothly with a strong arc light and welding splash interference and that tracking error can reach ±0.5, which is sufficient to satisfy actual welding requirements. The advantages of our algorithm are validated through several comparative experiments.