Global data association for multi-object tracking using network flows

Abstract

We propose a network flow based optimization method for data association needed for multiple object tracking. The maximum-a-posteriori (MAP) data association problem is mapped into a cost-flow network with a non-overlap constraint on trajectories. The optimal data association is found by a min-cost flow algorithm in the network. The network is augmented to include an Explicit Occlusion Model(EOM) to track with long-term inter-object occlusions. A solution to the EOM-based network is found by an iterative approach built upon the original algorithm. Initialization and termination of trajectories and potential

Multiple Target Tracking Using Frame Triplets

“…The best solution for K trajectories is found using a min-cost flow algorithm, while an outer loop searches for the optimal value K of number of objects. Berclaz et al [14] and Pirsiavash et al [6] show that the best set of paths and number of paths K in a min-cost flow network can be computed efficiently using a successive shortest paths algorithm, yielding the same solution as [5] but much more quickly. Although these approaches address the multi-frame data association problem in polynomial time, they do so by restricting the form of cost functions used to measure the quality of a trajectory to be a summation over only unary and pairwise edge weights.…”

“…Recent approaches formulate the multi-frame data association problem globally as a polynomial-time network flow problem [5,6,14]. Zhang et al [5] create a network with two nodes for each detection, a link between these nodes weighted by the probability that the detection is part of the solution, a link between detections in adjacent frames weighted by the cost of them being part of the same track, and a flow conservation constraint to ensure that no two trajectories share any observation.…”

“…Zhang et al [5] create a network with two nodes for each detection, a link between these nodes weighted by the probability that the detection is part of the solution, a link between detections in adjacent frames weighted by the cost of them being part of the same track, and a flow conservation constraint to ensure that no two trajectories share any observation. The best solution for K trajectories is found using a min-cost flow algorithm, while an outer loop searches for the optimal value K of number of objects.…”

“…However, there still remains the problem of recovering full trajectories in scenes with longer term occlusions. For this purpose we use the min-cost network flow formulation of [5] to link tracklets across occlusion gaps. A flow network is constructed with tracklets as nodes and with a directed edge placed from node T i to node T j when tracklet T j begins after T i has ended, as shown in Fig.…”

“…This greedy approach does not work well for closely-spaced, interacting targets [4]. Recent methods have attempted to bring more global information to bear by creating a graph of all the observations within a sequence [5][6][7], or by using a hierarchical solution method [8]. However, these graph-based methods only consider pairwise connections between observations, and thus are not able to capture higher-order constraints such as constant velocity that describe object motion across three or more frames.…”