Visual multi-object tracking via bi-level association strategy within Air-Traffic-Control surveillance videos

Human controllers of Air-Traffic-Control (ATC) system is vital in ensuring flight safety and efficiency. In this paper, we strive to take advantage of visual tracking multi-targets within ATC surveillance videos to assist the human air traffic controllers making decisions more quickly and precisely. In this special case, it should be established as online real-time multi-object tracking (MOT) of a large number of homogeneous small-sized targets under heavy occlusions. To solve this challenging problem we propose a bi-level association strategy that combining the advantages of batch-based association and sequential-based association. As none of distinctive appearance hint could be used during batch-based association, we formulate linking detection responses as self-organized growth of hyperedges in hypergraph. Extensive experiments have proven that our method outperforms the state-of-art within our specialized ATC surveillance videos.