Robust online multi-person tracking requires the correct associations ofonline detection responses with existing trajectories. We address this problemby developing a novel appearance modeling approach to provide accurateappearance affinities to guide data association. In contrast to most existingalgorithms that only consider the spatial structure of human appearances, weexploit the temporal dynamic characteristics within temporal appearancesequences to discriminate different persons. The temporal dynamic makes asufficient complement to the spatial structure of varying appearances in thefeature space, which significantly improves the affinity measurement betweentrajectories and detections. We propose a feature selection algorithm todescribe the appearance variations with mid-level semantic features, anddemonstrate its usefulness in terms of temporal dynamic appearance modeling.Moreover, the appearance model is learned incrementally by alternativelyevaluating newly-observed appearances and adjusting the model parameters to besuitable for online tracking. Reliable tracking of multiple persons in complexscenes is achieved by incorporating the learned model into an onlinetracking-by-detection framework. Our experiments on the challenging benchmarkMOTChallenge 2015 demonstrate that our method outperforms the state-of-the-artmulti-person tracking algorithms.
展开▼
