Graduated Moving Window Optimization as a Flexible Framework for Multi-Object Tracking

Continuous optimization methods for multiple object tracking allow to jointly estimate continuous object trajectories and perform implicit data association. However, the local minima that arise from including data association in a continuous optimization problem pose challenges. In addition, optimization is usually performed either over a fixed or an indefinitely growing time frame. This either discards valuable past information or is computationally unsustainable. Hence, in this work, a flexible continuous optimization based framework for multiple object tracking that accounts for these issues is proposed. The framework provides a unified approach to not only include data association, but also multiple motion models and temporary interactions between objects in a continuous optimization problem. It leverages the concept of graduated optimization, a heuristic, which allows avoiding local minima. The proposed framework's performance is benchmarked on a synthetic dataset, showing its capabilities and indicating areas of possible improvement.