GM-PHD filters for multi-object tracking in uncalibrated aerial videos

This paper presents an original answer to the difficult problem of multitarget detection and tracking with aerial images in an urban context. This framework multiplies difficulties for object detection as well as for tracking. Concerning detection, we have to deal with camera motion, strong parallax effects in urban areas, low image resolution, and unknown camera parameters. We propose to use an homographic transformation to compensate the camera motion and to combine geometric and intensity-based criteria for object detection. Tracking is also quite difficult, because it brings together all issues of multitarget tracking: extended targets with a probability of non-detection, hidden zones where no detection can be obtained, unresolved targets and an environment with a high false alarm rate. In this context, the recent probability hypothesis density (PHD) filter seems adapted and is used here in the Gaussian mixture (GMPHD) and in the Gaussian mixture cardinalized (GMCPHD) versions. Combining these two approaches leads to promising results on real urban sequences.