科研之友
微信
新浪微博
Facebook
分享链接
摘要
Structure from Motion (SFM) is a technique which takes image sequences or video sequences as input, then produces 3D point-cloud model. SFM technique has received a significant amount of attention in the past years both in computer vision and computer graphics communities motivated by applications in 3D reconstruction, augmented reality, virtual reality, map navigation, 3D change detection, scene completion, 3D salient object detection, and driverless car, due to the model produced by SFM technique has high geometric consistency with the scenes of images and video sequences. Generally speaking, feature tracking is a key fundamental component for SFM technique, which is used to produce feature matches from consecutive two image sequences or video frames, and the quality of the produced feature matches has a significant effect to the shape of the 3D model produced by SFM technique. Thus, in order to improve the geometric consistency of the 3D model with real scenes, a large number of feature tracking methods have been proposed, according to the different inputs, these methods can be roughly divided into two categories: the first type is used to handle unordered images where the resolutions are often not same to each other, and another one is used to process video sequences where each frame has the same resolution. However, among these feature tracking methods, most of them mainly focus on the problem that how to improve the accuracy and efficiency for specific applications such as indoor scene reconstruction and outdoor reconstruction where many repetitive features are existed, thus the essential drawbacks about the feature tracking problem has received few attentions. Although, some researchers have been proposed many state-of-the-art feature tracking methods in recent years, but it is very difficult to select a good one from the existing feature tracking methods for some special applications. In other words, we urgently need some guides to help us to make decision when selecting a suitable feature tracking method. To advance the research progress of feature tracking method, and improve the quality of the 3D model produced by SFM technique, in this paper we have made a comprehensive investigation for feature tracking methods. In detail, we first analyze some state-of-the-art feature tracking methods, and elaborate the core idea, advantages and disadvantages behind them. Second, we collect a significant amount of available resources such as feature detectors, feature descriptors and feature matching methods as well as benchmark datasets;these resources are very valuable both in theoretical research and practical applications. Third, we have made a comprehensive assessment for the state-of-the-art methods on various datasets such as 3D reconstruction dataset with scale variance and lighting variance, based on this evaluation we presented many valuable experiences which can be used as a guide to select a suitable feature tracking method from the existing methods for the specific applications. Fourth, we concluded some hard problems which should be solved in the future, and discussed what could affect the efficiency and accuracy of the feature tracking methods, this discussion can further push the research progress of feature tracking. Finally, we discussed the development trends of feature tracking, so as to point out the directions for feature research.
- 出版日期2018
- 单位合肥工业大学
