DuoCounter: An Approach for Counting Fish Using Tracking

Counting fingerlings remains a major challenge for fish farmers, as it demands both accuracy and efficiency to minimize errors and production costs. Automating this process through computer vision enables continuous monitoring and prevents duplicate counts by tracking individual fish. However, existing automated approaches that rely on delimiting a specific counting zone often suffer from identity switches. In many cases, fish are either missed or lost when passing through the zone and later re-identified after exiting and re-entering, leading to inaccuracies. To address these limitations, we propose DuoCounter, a tracking-based approach that integrates two reference lines to improve counting accuracy while maintaining a list of tracked objects and their corresponding keypoint-based features. These features are extracted and compared to resolve identity switches when they occur. The method was evaluated on real-world aquaculture videos with manually annotated ground truth, achieving a strong correlation with manual counts (0.97), as well as recall and precision scores of 0.91 and 0.83, respectively, using the YOLOv12 detector. The results demonstrate that DuoCounter effectively reduces identity switches, providing robust tracking and counting performance that matches manual counting while outperforming state-of-the-art methods