Persistent Tracking for Wide Area Aerial Surveillance
This page provides additional information about the conference publication below.
Data
The dataset WPAFB2009 from the U.S. Air Force Research Lab (AFRL) was used for evaluation. This dataset is publicly available, but cannot be redistributed. Please order the dataset from AFRL if you are interested.
Figure 1 shows the sequence that was used from the dataset. It is 1025 frames long, and 1408x1408 in size. It is also georeferenced, covering a region of about 429 m x 429 m. There are 410 tracks in total, with 24 targets in every frame on average. If you would like to compare with our approach, and have the dataset from AFRL, contact us at info@jgradient.com to get a more precise description.
Results
A noisy detector based on background subtraction was used to provide detections to the tracker. The performance of this detector is shown in Figure 2. Detection performance of the tracker is shown in Figure 3. Note that most of the false detects coming from the detector have been eliminated, while most of the correct detections have been successfully validated.
Tracking performance is shown in Figure 4. There are several examples of the tracker holding on to a target after it makes a complete stop, though some failures remain. The number of ID-switches and track breaks is lower than that of the competing approaches.
Reference
J. Prokaj , G. Medioni. Persistent Tracking for Wide Area Aerial Surveillance. In IEEE Conference on Computer Vision and Pattern Recognition , pages 1186-1193, 2014.