Automated tracking of marine life during fishing operations can provide quantitative data on animal movement, improving our understanding of animal behavior, and informing sustainable fishing gear design and use. While many fields use multi-object tracking (MOT) to automatically track objects in videos, its application remains limited in commercial fishing due to the challenges of collecting and automatically annotating footage of dynamic marine environments. We present the first application of MOT in a commercial Alaska walleye pollock ( Gadus chalcogrammus ) trawl, evaluating the feasibility of accurately tracking bycaught Pacific salmon ( Oncorhynchus spp.). We provide a detailed comparison of five sizes of YOLO12 detection models and four tracking algorithms: BoT-SORT, ByteTrack, Intersection over Union, and Centroid. Due to limited computational resources available on many commercial fishing vessels, we also evaluate tracker performance at low frame rates. Our findings demonstrate the feasibility of MOT in the commercial trawl environment, provide actionable conclusions about low-frame-rate tracking, outline effective tracker optimization techniques, and identify remaining challenges to improving MOT accuracy in this domain. The most accurate detection model (YOLO12x) and tracker (BoT-SORT) achieved a higher order tracking accuracy (HOTA) score of 59.5 on our challenging dataset. Evaluating tracker performance across frame rates revealed that simpler trackers outperformed more advanced Kalman-filter based trackers at low frame rates (10 and 7.5 FPS), indicating that frame rate should be considered during tracker selection. Object detection errors were the largest source of tracking error and should be the primary focus for improving tracking accuracy. To facilitate future MOT applications and research in this domain, our code, model weights, and dataset are publicly available. • Analyzed capabilities of MOT trackers in a commercial trawl environment. • Simple trackers can out-perform state-of-the-art trackers at low frame rates. • Detailed performance analysis of open-source object detectors and trackers on a novel dataset. • Outline methods for automated tracker optimization to new frame rates and domains.
Lurbur et al. (Sun,) studied this question.