Trajectory Reconnection

Problem Description

As identified by Paul St-Aubin, there are three cases of split or lost trajectories:

1. objects disrupted because of foregroung occlusion, typically by static foreground objects
2. objects tracked by different cameras, overlapping or not
3. objects stopping and lost by the tracker (in the same camera view)

Potential Methods

Reconnection methods:

1. measure similarity of trajectory end and beginning, stored in a bipartite graph, then run an optimization matching algorithm such as the Hungarian algorithm
   • similarity: position, time, predicted position, volume, type (if classified), appearance
2. for features lost because of foreground occlusion (one camera view), they could be searched by optical flow over a set of frames (and interpolated) (the quality control checks should be done through linear prediction).

Implementation Ideas

Regarding the method to store the result of reconnecting trajectories, it depends if the objects are initially stored in the same database or not.

• if stored in the same database, an option is to connect the features of the two objects to the same object.
  • if there is a time interval during which there is no feature, an artificial feature corresponding to the interpolated object position should be created and added to the feature group
• if stored in different databases, an option would be to merge both in a new database, with renumbering of objects and features

Potentially interesting papers

• Swaminathan Sankararaman and Pankaj K. Agarwal and Thomas Mølhave and Jiangwei Pan and Arnold P. Boedihardjo (2013). Model-driven matching and segmentation of trajectories. Proceedings of the 21st ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems - SIGSPATIAL'13 . Association for Computing Machinery (ACM).
• Sheikh, Y. and Shah, M. (2005). Object tracking across multiple independently moving airborne cameras. In Computer Vision, ICCV 2005. 2, 1555-1560. IEEE.