In this report, three popular methods for multi-pedestrian tracking are extended to a multi-category setting and tested on a large drone-based dataset. A thorough comparison of the algorithms is presented and a common shortcoming is identified. Building on this, a new tracking-by-detection based approach is developed that outperforms the other methods by a large margin. In addition, a state-of-the-art object detection model is adapted for the drone imagery, since no public detections are available for the dataset.
