Re-identification

Youssef Alami Mejjati       Marc Decombas.

Person re-identification is an active open subject that has for the moment being study for video protection with different cameras in a short period (Constant appearance), and closed set (person to re-identify is already available in the dataset). It can be defined as the task of recognizing a person through one or several cameras. It is one of the most important challenging tasks in video surveillance, due to the fact that lots of different algorithms are necessary. People detection, tracking, feature extraction, descriptor generation and measurement of the similarity. Moreover, when someone is tracked through one or several cameras, lots of difficulties like low resolution images, low frame rates, difference of luminosity/contrast, same subject seen by differents angles, occlusion and shape changing have to be outfaced. In this work, we are trying to deal with a real case, meaning that we are on an open set problem. When a person arrives, it is added in the current dataset, and when one leaves, it should be re-identified to be suppressed from the dataset. The dataset is changing in the time and error of re-identification will greatly influence the performance by keeping person that is not here anymore in the dataset. In this paper, after describing and highlighting the strengths and the weaknesses of the different process that have knowledge, this work is one of the first trying to deal with an openset problem. 1) Traditional dataset will be adapted to manage the open set problem and 2)A new evaluation will also be presented. 3) A new method based on salience learning and applied not on one image but group of images is presented. Comparison is done with the state of the art.