Learning Optimal Parameters For Multi-target Tracking

Multi-target tracking problems are traditionally tackled in two different ways. One way is to first group detections into candidate tracklets and then perform scoring and association of these tracklets [5, 6], this can be done in either an online/streaming fashion or an offline/batch fashion and it allows tracklets to be scored with richer trajectory and appearance models. Another approach is to attempt to include higher-order constraints directly in a combinatorial framework [1, 2]. In either case, there are a large number of parameters associated with these richer models which become increasingly difficult to set by hand and necessitate the application of machine learning techniques. In this paper, we describe an end-to-end framework for learning parameters of min-cost flow multi-target tracking problem with quadratic trajectory interactions including suppression of overlapping tracks and contextual cues about co-occurrence of different objects. Our approach utilizes structured prediction with a tracking-specific loss function to learn the complete set of model parameters. Under our learning framework, we evaluate two different approaches to finding an optimal set of tracks under quadratic model objective based on an LP relaxation and a novel greedy extension to dynamic programming that handles pairwise interactions. In a min-cost flow multi-target tracking problem, the set of optimal (most probable) tracks can be found by solving an integer linear program (ILP) over flow variables f.