Section: New Results

Global Tracking of Multiples Actors

Participants : Julien Badie, François Brémond.

We propose a new approach for long term tracking of individuals. Our main objective is to design a tracking algorithm for people reidentification [30] that can track people even if they come back in the scene after leaving it. This algorithm is based on covariance matrix and we have also added some contextual information of the scene (for instance, zones where people can enter the scene) to improve tracking performance. In addition, a basic noise detection system and a tracking correction system are proposed in order to handle short-term tracking errors such as multiplication of IDs corresponding to only one individual. The noise detection system is designed to find and remove objects that are detected in a very small number of consecutive frames (for instance 4) and disappear afterward. The tracking correction system associates IDs recently lost with IDs that have just started to be tracked based on geometrical features and 3D distance criteria.

As a result, the tracking quality is significantly improved on 5 video sequences tested from the ETISEO dataset (http://www-sop.inria.fr/orion/ETISEO/ ). The people reidentification algorithm gives encouraging results for future work. The number of IDs associated to one individual is reduced (on average $50\%$ less) and the tracking quality improves due to the IDs stability. This algorithm can detect not only people re-entering the scene but also trajectory interruptions due to occlusions or misdetections.

This approach could enable the detection of new kinds of events on video sequences such as long range people tracking on a camera network.

Figure 16. The green lines are the ground-truth and the blue lines are the tracking results using both reidentification and noise removal (left) and with only noise removal (right). The man is tracked throughout the whole video with a single ID (68).