Scholarlink

<p>Marine fish tracking is critical for sustainable fishery management; however, this task remains challenging due to visual similarities among conspecific fish, nonlinear motion patterns, and frequent occlusions. To address these challenges, we propose Marine Multi-Fish Tracking (MMFTrack), an efficient tracking model tailored for marine multi-fish scenarios, incorporating four key innovations: (1) a Fish-Xception-SPA appearance feature extractor that leverages depthwise separable convolutions and spatial attention mechanisms to enhance discriminative capability; (2) robust nonlinear motion modeling using an unscented Kalman filter (UKF); (3) a dynamic occlusion update strategy that adaptively adjusts the fusion weights between motion prediction and appearance observation based on the duration of occlusion; and (4) a category-gated association mechanism to prevent cross-species identity switches. Comprehensive evaluations on a proprietary South China Sea fish dataset demonstrate that MMFTrack achieves performance gains over the baseline StrongSORT model in HOTA (+2.8), MOTA (+0.7), and IDF1 (+7.9) while maintaining real-time performance at 83.3 FPS. Furthermore, rigorous generalization experiments on public benchmarks confirm the robustness of MMFTrack, reducing identity switches by 88.9% (4 vs. 36) on the FISHTRAC dataset and by 31.7% (1,618 vs. 2,369) on the SEAMAPD21 dataset. This study provides an efficient and robust solution for marine fish tracking, enabling deployment on marine robotic platforms for population abundance estimation and establishing a technical foundation for data-driven fishery governance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MMFTrack: an occlusion-aware and species-specific real-time tracking model for marine multi-fish

  • 期刊论文

  • 发表日期：

    2025年10月29日

查看全文
  • Jiaxin Li,
  • Hualing Li,
  • Jiaxin Huo,
  • Yonglai Zhang

摘要

Marine fish tracking is critical for sustainable fishery management; however, this task remains challenging due to visual similarities among conspecific fish, nonlinear motion patterns, and frequent occlusions. To address these challenges, we propose Marine Multi-Fish Tracking (MMFTrack), an efficient tracking model tailored for marine multi-fish scenarios, incorporating four key innovations: (1) a Fish-Xception-SPA appearance feature extractor that leverages depthwise separable convolutions and spatial attention mechanisms to enhance discriminative capability; (2) robust nonlinear motion modeling using an unscented Kalman filter (UKF); (3) a dynamic occlusion update strategy that adaptively adjusts the fusion weights between motion prediction and appearance observation based on the duration of occlusion; and (4) a category-gated association mechanism to prevent cross-species identity switches. Comprehensive evaluations on a proprietary South China Sea fish dataset demonstrate that MMFTrack achieves performance gains over the baseline StrongSORT model in HOTA (+2.8), MOTA (+0.7), and IDF1 (+7.9) while maintaining real-time performance at 83.3 FPS. Furthermore, rigorous generalization experiments on public benchmarks confirm the robustness of MMFTrack, reducing identity switches by 88.9% (4 vs. 36) on the FISHTRAC dataset and by 31.7% (1,618 vs. 2,369) on the SEAMAPD21 dataset. This study provides an efficient and robust solution for marine fish tracking, enabling deployment on marine robotic platforms for population abundance estimation and establishing a technical foundation for data-driven fishery governance.