Multi-Vehicle Detection and Tracking in Aerial Imagery Sequences using Deep Learning Algorithms

Kurzfassung

Multi-Object Tracking (MOT) describes the task of identifying all objects in an image and following them over a sequence of frames. A wide range of Deep Learning-based MOT methods has been developed for the autonomous driving domain, but not for aerial imagery. Applications, for example traffic analysis, depend on such tracking capability. This thesis addresses multi-vehicle tracking in top-down aerial imagery sequences. MOT algorithms of other domains are applied on an aerial dataset to create a benchmark. The best method is then identified by using a set of appropriate metrics. Based on that, a variety of modifications are performed to inspect the effect on the tracking capability. This variety consists of a vehicle orientation prediction, an object re-identification feature, and a motion prediction module. Experiments with different configurations are conducted and evaluated. A MOTA score of 78.5 is achieved on the KIT AIS vehicle dataset.