Multiple People Tracking

TNT members involved in this project:
Timo Kaiser
Prof. Dr.-Ing. Bodo Rosenhahn
Dr.-Ing. Dipl.-Math. Roberto Henschel
Prof. Dr.-Ing. Laura Leal-Taixé
Show active Staff only

Multiple people tracking is a challenging task of the computer vision domain, with applications in surveillance, action recognition and for example human-computer interaction systems.

Whereas single object tracking can be considered as almost solved, multiple people tracking still remains a difficult problem with ongoing research, due to the complex interaction between persons, that needs to be modeled in order to obtain good results.

We are interested in finding the global optimal solution to the data association problem. In order to do so, we propose to solve the tracking problem by using knowledge from graph theory:

  • We model the association problem as a hierarchical tracking problem in a DAG, which reduces wrong decisions in each step compared to other hierarchical approaches and we solve the association problem by efficiently computing a minimum cost arborescence.
  • We model the association problem in a network flow graph, include social and grouping behavior between objects and solve it by applying the simplex algorithm.


Show all publications
  • Andrea Hornakova*, Timo Kaiser*, Michal Rolinek, Bodo Rosenhahn, Paul Swoboda, Roberto Henschel, (* equal contribution)
    Making Higher Order MOT Scalable: An Efficient Approximate Solver for Lifted Disjoint Paths
    International Conference on Computer Vision (ICCV), IEEE, October 2021
  • Andrea Hornakova*, Timo Kaiser*, Bodo Rosenhahn, Paul Swoboda, Roberto Henschel, (* equal contribution)
    Higher Order Multiple Object Tracking for Crowded Scenes
    Computer Vision and Pattern Recognition Workshops (CVPRW) , June 2021
  • Andrea Hornakova*, Roberto Henschel*, Bodo Rosenhahn, Paul Swoboda, (* equal contribution)
    Lifted Disjoint Paths with Application in Multiple Object Tracking
    Proceedings of the 37th International Conference on Machine Learning (ICML), July 2020