Satellite Multi-Vehicle Tracking under Inconsistent Detection Conditions by Bilevel K-Shortest Paths Optimization

Tracking multiple vehicles in optical satellite videos can be achieved by detecting vehicles in individual frames and then link detections across frames. This is challenging since the low spatial resolution of satellite videos would lead to miss or partial detections. Failure in consistently detecting an identical vehicle would terminate a trajectory for loss of identity, or reinitialize a new trajectory for re-detection. Both would result in fragmented tracklets for an identical vehicle. In this paper, we propose a two-step global data association approach for multiple target tracking, which first generates local tracklets then merges them to global trajectories. In order to handle the frequent terminals of tracklets caused by the miss or partial detections, the spatial grid flow model has been extended to temporal neighboring grids to cover the possible connectivities in a wider temporal neighboring, so that the detection association could match temporal-unlinked detections. In the tracklet association model, we customize a tracklet transition probability based on Kalman Filter to link tracklets with large temporal intervals. A near-optimal solution to the proposed two-step association problem is found by Bilevel K-shortest Paths Optimization. Experiments on a satellite video show that our approach improves both detection and tracking performance of moving vehicles.