Tracking many vehicles in wide area aerial surveillance

Wide area aerial surveillance data has recently proliferated and increased the demand for multi-object tracking algorithms. However, the limited appearance information on every target creates much ambiguity in tracking and increases the difficulty of removing false target detections. In this work we propose to learn motion patterns in wide area scenes and take advantage of this additional information in tracking to remove false alarm and reduce tracking error. We extend an existing multi-object tracker for wide area imagery by incorporating the motion pattern data as further probabilistic evidence. Scalability is ensured by dividing the imagery into tiles, processing each tile in parallel, and handing off tracks between tiles when necessary. Evaluation on sequences from a real wide area imagery dataset shows this approach outperforms a competing tracker not making use of such data.