This project aims to develop methods for long term tracking of multiple objects in video sequences. Multiple object tracking has received the attention of the image processing community in the last 5 years, fostered by surveillance applications and by Model Based Video Coding (MPEG).
The first works addressed short-term tracking and recognition of activities. More recent works have tried to address long term tracking of moving objects. This is a more difficult problem since it involves the ability to disambiguate the trajectories of the objects after they were grouped and occluded for some time.
This project aims to address this problem. We wish to detect moving regions in video sequences and to develop algorithms to label each region in a consistent way along the whole video sequence. An additional difficulty concerns the presence of merged regions which can not be identified by a single label. Probabilistic models, namely probabilistic networks, will be adopted to perform this task and to propagate probable labelling scenarios. The tracking algorithms will be applied in the context of urban surveillance.