Multiple Object Tracking Using Local Motion Patterns

Abstract

This paper presents an algorithm for multiple-object tracking without using object detection. We concentrate on creating long-term trajectories for unknown moving objects by using a model-free tracking algorithm. Each individual object is tracked by modeling the temporal relationship between sequentially occurring local motion patterns. The algorithm is based on shape and motion descriptors of moving objects, obtained at two hierarchical levels from an event understanding system. By considering both local and global motion patterns, two sets of initial tracks, called linklets, are obtained. Then, a set of sparse tracks, referred to in the literature as tracklets, is produced by grouping linklets demonstrating similar motion patterns. This produces two sets of independent tracklets, referred to as the lowand high-level tracklets. We adopt Markov Chain Monte Carlo Data Association (MCMCDA) to estimate a varying number of trajectories given a set of tracklets as input. To this end, we formulate tracklet association as a Maximum A Posteriori (MAP) problem to create a chain of tracklets. The final output of the data association algorithm is a partition of the set of tracklets such that those belong to individual objects have been grouped. This yields individual tracks for each object in a video.