Metric definition and optimization for the appearance information in Multi-Object Tracking

Details

Supervisors

De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

Multiple Object Tracking, or MOT, is a computer vision tasks, which consists in locating detections of multiple objects in video, obtaining information about these objects and associating them into trajectories. This task has many applications, including autonomous driving, pedestrian and crowd monitoring, or player localization in a sport setting. One of the important steps in MOT is feature extraction, where appearance information are obtained on the targets in order to help determine their trajectories. Currently, there are no ways to evaluate the capacity of a given set of extracted appearance information to enable good tracking performances, other than performing tracking directly. In this thesis, we proposes a new metric which evaluates such a thing. Our new metric evaluates a ranking of the set of all possible pairs of detections, ordered by distance between the appearance information distance between each elements of the pairs. In addition, we propose a method for modifying the rankings of pairs of detections to improve the output of our new metric. For this, we used the concept of side information, which is the information outputted by a feature extraction model besides the appearance information and which can indicate us for example the certainty that we have that we can trust the appearance information. We tested our new metric on two datasets, and tried to improve these result using our ranking modification.