Abstract
Automated fish detection and tracking on recreational fishing vessels presents significant technical challenges due to motion blur, visual clutter, and frequent occlusions. This thesis proposes a hybrid fish tracking framework that combines state-of-the-art models for robust and precise detection and identity-aware tracking in real-world aquatic conditions. The system integrates YOLOv12 for real-time object detection, DeepSORT for assigning and maintaining consistent object identities across frames, and the Segment Anything Model (SAM2.1) for fine-grained, mask-based fish tracking. By linking DeepSORT’s identity assignment with SAM2.1’s mask propagation, the system enhances both spatial accuracy and temporal continuity, enabling effective fish monitoring over extended video sequences. Experimental evaluations demonstrate that this hybrid pipeline outperforms standalone models, particularly in visually complex environments commonly encountered in recreational fishing. The proposed approach is optimized for onboard efficiency and holds practical potential for automated logging, ecological monitoring, and event-based catch detection, reducing the need for manual oversight during fishing activities.