Tracking underwater fish species is essential for population monitoring, stock assessment, ecological research, and the protection of endangered species. Although image processing and computer vision techniques have advanced fish detection and classification, multi-object tracking in natural underwater environments remains significantly more challenging and relatively underexplored. This work utilizes the Gulf Fishery-Independent Survey of Habitat and Ecosystem Resources Dataset 2024 (GFISHERD24) dataset, collected from the Gulf of America in 2024, which contains frame-annotated videos of fish in their natural habitats. The dataset presents substantial real-world challenges, including occlusions caused by other fish or vertical cage structures, multiple fish per frame, low-light conditions, and variable visibility. To overcome these challenges, we employ YOLOX, an anchor-free, high-performance object detector, and adapt the ByteTrack algorithm for robust multi-object tracking. Unlike traditional trackers, ByteTrack retains both high and low confidence detections, allowing more accurate association of partially occluded or low-visibility fish. The GFISHERD24 dataset was converted to MS COCO format for detection training and to MOT17 format for tracking evaluation. On the test set, the adapted ByteTrack achieved a good performance; however, a major challenge in our dataset arises from fish moving behind vertical cage bars, causing significant visual obstruction. These occlusions often break a fish’s existing track and result in new track IDs, ultimately distorting fish counts. To mitigate this, we removed the vertical-bar obstructions from image frames and retrained and tested the models using the refined dataset. This preprocessing noticeably increased MOTA and reduced ID switches. In conclusion, with improved image frame preprocessing to handle structural occlusions, the adapted ByteTrack framework demonstrates strong effectiveness for large-scale underwater fish tracking in challenging, real-world environments.