Fast and accurate pupil tracking, even in environments with limited computing resources, is critical for applications such as eye tracking and driver drowsiness warning systems. This paper proposes BF-YOLOv3-tiny for fast and accurate pupil tracking. ...
Fast and accurate pupil tracking, even in environments with limited computing resources, is critical for applications such as eye tracking and driver drowsiness warning systems. This paper proposes BF-YOLOv3-tiny for fast and accurate pupil tracking. Key improvements include: A bi-directional fusion method was applied to interconnect low-resolution and high-resolution feature maps, and anchors boxes were selected by considering distribution changes due to data augmentation during training process. In addition, a signal processing technique to remove grid sensitivity and an IoU-based loss function were adopted when model predicts the bounding boxes. Data provided by Department of Ophthalmology of Pusan National University hospital was used to evaluate the proposed model, and the results were compared and analyzed through comparative experiments with five lightweight networks.
The proposed model shows performance up to 98.0 AP 50, 78.8 AP 75, and 44.6 AP T, outperforming compared to existing YOLOv3-tiny and other lightweight networks. Lastly, as a result of implementing the model with the best performance on NVIDIA Jetson Nano, it achieved up to 100.0 AP 50 and 26.2 FPS, demonstrating its feasibility and an accurate and real-time pupil tracking system even in an environment with limited computing resources.