Multi-camera setups are used in scenarios like visual surveillance, 3D reconstruction, human modelling, virtual reality, etc. Despite the fact that a lot of research has been done on single camera people detection and tracking, still this approach remains restricted in cases e.g. in identifying small scales, occlusions; including feature and dataset based limitations. Multi-camera setups inherently provide more information for the under constrained scenario of video detection and tracking. The context of this internship is to perform people detection and/or tracking using multiple overlapping cameras in order to perform efficient automated visual surveillance.
Recent notable works for multi-camera people tracking in overlapping scenarios include homography based techniques and probabilistic approaches. At first, the student is expected to study these approaches in order to work with the techniques and library of tools developed at LEOST, IFSTTAR. The main goal of this work is to contribute towards improving people detection and/or tracking efficiencies using the 3D geometric models and/or probabilistic techniques, working with datasets such PETS 2009 or the datasets developped in the BOSS Celtic project, by including the research ideas proposed by the student him/herself. The first idea and way of improvement will be to take into account the dynamic of the objects in the scene to better separate the objects. The second idea could be to use dictionary learning techniques to improve sparsity of the representation and to represent one people in several configuration by several atoms.
This internship provides an opportunity for both research and programming experience.
Student Profile: 2nd year masters or 3rd year engineering student in Computer Sciences, Applied Mathematics or Image Processing.
Skills Required: MATLAB and/or C/C++. Prior knowledge in 3D Computer Vision, Machine Learning will be considered as an additional plus point.
Application: Send an application with CV, letter of motivation and master (1&2) transcripts by email to the following contact. Students are encouraged to email in case of any inquiry about the project or the internship.
Sébastien AmbellouisEmail: firstname.lastname@example.org
LEOST, IFSTTAR, Lille - Villeneuve d'Ascq (FRANCE)
1. Fleuret, F., Berclaz, J., Lengagne, R., and Fua, P. (2007). P.: Multi-camera people tracking with a probabilistic occupancy map. IEEE Transactions on Pattern Analysis and Machine Intelligence.
2. Ren, J., Xu, M., and Smith, J. S. (2012). Pruning phantom detections from multiview foreground intersection. In ICIP, pages 1025–1028.
3. Utasi, A. and Benedek, C.(2012). A bayesian approach on people localization in multi-camera systems. IEEE Transactions on Circuits and Systems for Video Technology.
4. M. Liem and D. M. Gavrila (2013). A comparative study on multi-person tracking using overlapping cameras. Lecture Notes in Computer Science: Proc. of the International Conference on Computer Vision Systems (St.Petersburg, Russia), vol. 7963
5. Owais Mehmood, Sébastien Ambellouis, and Catherine Achard. Launch these manhunts! shaping the synergy maps for multi-camera detection. In International Conference on Computer Vision Theory and Applications (VISAPP), volume 2, pages 528-535
6. Owais Mehmood, Sébastien Ambellouis, and Catherine Achard . Exploiting 3d geometric primitives for multicamera pedestrian detection. In International Conference on Advanced Video and Signal based Surveillance (AVSS)
(c) GdR 720 ISIS - CNRS - 2011-2015.