Deep network flow for multi-object tracking
View Patent ↗A multi-object tracking system and method are provided. The multi-object tracking system includes at least one camera configured to capture a set of input images of a set of objects to be tracked. The multi-object tracking system further includes a memory storing a learning model configured to perform multi-object tracking by jointly learning arbitrarily parameterized and differentiable cost functions for all variables in a linear program that associates object detections with bounding boxes to form trajectories. The multi-object tracking system also includes a processor configured to (i) detect the objects and track locations of the objects by applying the learning model to the set of input images in a multi-object tracking task, and (ii), provide a listing of the objects and the locations of the objects for the multi-object tracking task. A bi-level optimization is used to minimize a loss defined on a solution of the linear program.
1. A multi-object tracking system, comprising:
at least one camera configured to capture a set of input images of a set of objects to be tracked;
a memory storing a learning model configured to perform multi-object tracking by jointly learning arbitrarily parameterized and differentiable cost functions for all variables in a linear program that associates object detections with bounding boxes to form trajectories; and
a processor configured to (i) detect the objects and track locations of the objects by applying the learning model to the set of input images in a multi-object tracking task, and (ii), provide a listing of the objects and the locations of the objects for the multi-object tracking task,
wherein a bi-level optimization is used to minimize a loss defined on a solution of the linear program;
wherein the processor is further configured to detect the objects and track the locations of the objects by solving a Linear Programming (LP) problem for any of the input images inside a sliding window Q, moving the sliding window Q by Δ frames, and solving a resultant new LP problem, where 0<Δ<W ensures a minimal overlap between a LP solution for the LP problem and another LP solution for the resultant new LP problem.
2. The multi-object tracking system of claim 1 , wherein the LP solutions comprises different sets of trajectories, and the processor is further configured to determine a matching cost for each of the trajectories that is inversely proportional to a number of shared detections there between.