One-Shot Learning of Scene Locations via Feature Trajectory Transfer

Author(s):  
Roland Kwitt ◽  
Sebastian Hegenbart ◽  
Marc Niethammer
Keyword(s):  

2004 ◽  
Author(s):  
Yasuhiro Minami ◽  
Erik McDermott ◽  
Atsushi Nakamura ◽  
Shigeru Katagiri






Author(s):  
C. K. Liu ◽  
K. Qin ◽  
K. Chen ◽  
R. Ma

Abstract. Constrained by road network structure, travel choice and city function zoning, GPS trajectory data exhibits significant spatiotemporal correlation. Unveiling the clustering and distribution patterns of GPS trajectory can help to better understand the travel behaviour as well as the corresponding spatial and temporal characteristics. This paper proposes an approach to identify and visualize the aggregation pattern from GPS trajectory data. Firstly, slow feature trajectory sequences are extracted from raw taxi trajectory data. Together with taxi states information, these sequences are processed as shorter length tracks for faster discovery of cluster similarity. Thereafter, the temporal and spatial similarity and dissimilarity metrics between the trajectories are established, and the temporal and spatial distances between the trajectories are defined to form a space-time cylinder model. Next, based on the idea of density clustering, the DBSCAN spatiotemporal expansion of trajectory data is proposed. Feature trajectory sequences are then clustered into groups with high similarity. Finally, for a more intuitive understanding of the trajectory aggregate distribution, time dimension info of each point in the sequences is used as Z axis, thus the sequences are stretched on the map in different colour for 3D visualization. The proposed method is validated by a case study of taxi trajectory data analysis in Wuhan City, China.



2013 ◽  
Vol 23 (12) ◽  
pp. 2105-2115 ◽  
Author(s):  
Yang-Ho Cho ◽  
Ho-Young Lee ◽  
Du-Sik Park


2014 ◽  
Vol 519-520 ◽  
pp. 640-643 ◽  
Author(s):  
Jing Dong ◽  
Yang Xia

In this paper, a real-time video stabilization algorithm based on smoothing feature trajectories is proposed. For each input frame, our approach generates multiple feature trajectories by performing inter-frame template match and optical flow. A Kalman filter is then performed to smooth these feature trajectories. Finally, at the stage of image composition, the motion consistency of the feature trajectory is considered for achieving a visually plausible stabilized video. The proposed method can offer real-time video stabilization and its removed the delays for caching coming images. Experiments show that our approach can offer real-time stabilizing for videos with various complicated scenes.



Sign in / Sign up

Export Citation Format

Share Document