Abstract: The performance of joint feature extraction is degraded greatly during welding because of severe disturbances,so a prediction-matching-estimation combined close-looped iterated method is proposed for weld feature tracking.The position and shape of the joint stripe profile in the image is predicted at first,then the joint is recognized by a model-based profile matching method,and lastly the weld trajectory is filtered and estimated based on the confidence of the extracted feature.Experiments show that the method promotes the joint recognition results to ensure a robust and real-time weld seam tracking even in the face of gross error or failure occurrence in joint feature extraction.