Accurate natural contour tracking for non-rigid object?

The natural contour extraction during non-rigid object tracking is a challenging task in computer vision. Most tracking-by-detection methods are based on rectangular bounding-boxes, and this leads to compounding tracking errors in subsequent frames. This paper present an accurate natural contour tracking method for non-rigid object in video, there are three main contributions. Firstly, we combined a real-time superpixel segmentation technique with natural contour tracking task to reduce the computational cost while providing very favorable boundary structural information. Secondly, we proposed an object-oriented natural contour extraction method for non-rigid objects. Thirdly, we propose a saliency-based natural contour tracker for non-rigid object. The proposed method can effectively handle the tracking problems introduced by foreground interference, complex background and severe changes in shape, scale as well as illumination. Therefore, our method is able to track the non-rigid target object robustly and provide an accurate natural contour. Our experimental results on several publicly available datasets show that our method outperforms some state-of-the-art non-rigid object tracking approaches both qualitatively and quantitatively.