| dc.description.abstract | This master's thesis focuses on the evaluation and exploration of detection and tracking algorithms for fish in a dense underwater environment. The primary objectives were to achieve precise and accurate fish detection and to track fish over an extended period. The thesis explores the performance of two object detection algorithms, YOLOv4 and YOLOv8, as well as their integration with the DeepSORT tracking algorithm. The algorithms were trained and evaluated using a dataset collected from a densely populated underwater fish tank. The dataset was manually annotated using bounding box annotation techniques to accurately label the objects of interest.
The results demonstrated the effectiveness of both YOLOv4 and YOLOv8 in detecting fish in densely populated environments. However, YOLOv8 achieved a significantly higher mAP50-95 score, indicating better localization and detection accuracy. It proved more adept at precisely locating the position of detected fish, leading to improved overall detection performance.
In terms of fish tracking the combination of DeepSORT and YOLOv8 showed the best overall performance, as evidenced by higher MOTA and IDF1 scores, and lower MOTP scores. However, tracking individual fish over extended periods presented challenges due to occlusions and rapid trajectory changes, leading to a high number of identity switches.
By evaluating and exploring the effectiveness of detection and tracking algorithms, this thesis contributes to the advancement of fish monitoring techniques in aquaculture. The findings provide valuable insights into the performance of YOLOv4 and YOLOv8 and the potential of DeepSORT for accurate and reliable fish detection and tracking. The results and methodologies presented in this study lay the groundwork for further research and development in the field, aiming to enhance fish welfare, optimize resource management, and improve efficiency in aquaculture practices. | |
