Department of Computer Engineering
Keypoint Estimation Based Multi-Object Tracking in Aerial Images
Fikret Efe Doğan
(Supervisor: Asst. Prof. Ercüment Çiçek)
Computer Engineering Department
Multi-object tracking is one of the trending research areas in computer vision, where the goal is to assign an identity to each object in a given image sequence. Current deep learning architectures that involve tracking by detection paradigm suffer from a high number of trainable parameters, long training sessions, and low frame-per-second (fps) rates. Furthermore, most networks fail to track consistently in aerial images containing small objects. Since the performance of multi-object tracking is heavily dependent on the object detection accuracy, we discard the predefined anchor placements from the backbone and adopt a keypoint detection mechanism to implement an end-to-end trainable convolutional neural network. This also helps to narrow down the hyperparameter search during experiments. The network is capable of producing tracking predictions at one go by temporally associating the detected keypoints.
DATE: 08 November 2021, Monday @ 15:50 Zoom