Real-time interactive multi-target tracking using kernel-based trackers

Although kernel-based methods have been demonstrated effectively in solving single-target tracking problem, facing more complicated multi-target tracking task, most of them still suffer from `singularity´ problem because they usually simply track each target independently ignoring their interactions and thus fail when occlusions occur or distracters appear. In this paper, we discuss a very common scenario of multi-target tracking, in which a moving object has its own destination in a short time and pays attention on some of its neighboring objects. This phenomenon exists in many practical tracking applications such as human tracking, traffic monitoring, video surveillance, etc., where an object usually moves towards a particular direction but meanwhile detours when close to others to avoid potential collision. In particular, we propose two novel interaction models and embed them into a kernel-based tracking framework to derive our interactive kernel-based multi-target tracker. Experimental results on challenging videos demonstrate its superiority.