Struck is a structured output tracking framework that improves adaptive visual object tracking by directly predicting the change in object configuration between frames. Unlike traditional methods that separate tracking and learning, Struck integrates these processes, avoiding the need for intermediate classification steps. It uses a kernelised structured output support vector machine (SVM) learned online for adaptive tracking. To maintain high frame rates, a budgeting mechanism limits the number of support vectors, preventing unbounded growth. The framework also supports GPU-based tracking, enabling efficient real-time performance. Experiments show that Struck outperforms state-of-the-art trackers on benchmark videos and allows easy incorporation of additional features and kernels, enhancing tracking performance. The paper introduces a novel approach to structured output prediction, addressing issues in traditional tracking-by-detection methods by directly linking learning to tracking and avoiding artificial binarisation steps. The framework is implemented on both CPU and GPU, with the GPU version achieving high frame rates. The paper also discusses practical considerations, including parameter vectors, kernel functions, and multiple-kernel learning, which contribute to the effectiveness of the tracking algorithm.Struck is a structured output tracking framework that improves adaptive visual object tracking by directly predicting the change in object configuration between frames. Unlike traditional methods that separate tracking and learning, Struck integrates these processes, avoiding the need for intermediate classification steps. It uses a kernelised structured output support vector machine (SVM) learned online for adaptive tracking. To maintain high frame rates, a budgeting mechanism limits the number of support vectors, preventing unbounded growth. The framework also supports GPU-based tracking, enabling efficient real-time performance. Experiments show that Struck outperforms state-of-the-art trackers on benchmark videos and allows easy incorporation of additional features and kernels, enhancing tracking performance. The paper introduces a novel approach to structured output prediction, addressing issues in traditional tracking-by-detection methods by directly linking learning to tracking and avoiding artificial binarisation steps. The framework is implemented on both CPU and GPU, with the GPU version achieving high frame rates. The paper also discusses practical considerations, including parameter vectors, kernel functions, and multiple-kernel learning, which contribute to the effectiveness of the tracking algorithm.