Abstract

We introduce an online approach to learn possible ele- mentary groups(groups that contain only Iwo targets)for inferring high level context that can be used to improve multi-target tracking in a data-association based frame- work.Unlike most existing association-based tracking ap- proaches that use only low level information(e.g.,time,ap- pearance,and motion)to build thte affiniry model and con- sider each target as an independent agent,we online learn social grouping behavior to provide addirional information for producing more robust tracklets affinities.Social grorp- ing behavior of pairwise targets is first learned from con- fident tracklets and encoded in a disjoint grouping graph. The grouping graph is fiurther completed with the help of group rracking.The proposed method is efficien,handles group merge and split,and can be easily integrated into any basic affiniry model.We evaluate our approach on two ptb- lic datasets,and show significant improvements compared with state-of-the-art methods.