Generic Object Crowd Tracking by Multi-Task Learning

Multiple Object Tracking (MOT) is an important problem for its various applications. In general, approaches for MOT can be categorised into two types, sequential ones and batch ones. Sequential ones utilise observations from frames up to the time while batch ones use observations from all frames of a video. For the significant progresses achieved in the pedestrian detection field, most existing work for MOT follows the batch fashion (i.e. it employs a pedestrian detector to carry out detection in each frame in advance, and then handles MOT as a data association problem by treating the detection responses of all the frames as observations). However, little attention has been paid to detection and tracking of multiple objects of an arbitrary type. In this paper, we tackle the problem of tracking multiple objects without limitation of the type of the objects. Furthermore, we show how this problem can be formulated within the Multiple Task Learning (MTL) framework [3].