Solving Multiple People Tracking in a Minimum Cost Arborescence

For many applications of computer vision, it is necessary to localize and track humans that appear in a video sequence. Multiple people tracking has thus evolved as an ongoing research topic in the computer vision domain. A commonly used approach to solve the data association problem within the tracking task is to apply a hierarchical tracklet framework. Although there has been great progress in such a model, mainly due to its good bootstrapping capabilities, so far little attention has been drawn to improve the quality of the tracklets itself. A main issue of the hierarchical frameworks, as used in the common literature, is that they make hard decisions at each iteration of the association step. Especially in ambiguous situations, tracklets are still being merged or removed so that the system is prone to error propagation. To avoid these problems, we propose a new framework where unreliable decisions are prevented. Instead, unclear aggregations are being postponed to a later iteration, when more information is available. To maintain the possible associations of tracklets in difficult situations, we propose a new trajectory model, which we call tree tracklets. While recent multiple people trackers model the association problem mainly in a flow network, we employ a rooted, directed and weighted graph G = (V, E, w) which is of a simpler structure, in particular has fewer nodes and edges. Thereby, we obtain the global-optimal solution of each iteration in linear time in the number of nodes by computing a minimum cost arborescence.