Tracking Planar Objects by Segment Pixels
Zhongguan Zhai, Shang Sun, Junjie Liu · 2021 3rd International Academic Exchange Conference on Science and Technology Innovation (IAECST) · 2021
The methods of representing a planar object by four corner points have been widely used in planar object tracking. These previous approaches match tracking objects between video frames by handcraft features or by directly minimizing errors. However, these methods lack the ability to classify objects and backgrounds with high-level semantic information, resulting in a less robust performance in complex scenes such as motion blur. In this paper, we propose a robust approach to pixel-level tracking by learning semantic information about the tracking object. We solve the planar object tracking in terms of finding each pixel in the region where the planar object is located. Our proposed STMPOT network can memorize the object and background in-formation of each frame. The proposed STMPOT network learns to classify the pixels belonging to the object or background of the current frame. We transform the POT dataset into a POT-seg dataset to train our method. We also compare our method with other planar object trackers by the proposed evaluation metrics, and the experimental results show that our method achieves the best results in most scenes.