.png){kind=icon}
.png){kind=icon}
Return
DA-MFT: Unsupervised Multiple Fish Tracking Framework Based on Domain Adaptation
.png){kind=icon}
PRE
AI
DOI：10.1109/TII.2025.3621012.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multiple fish tracking (MFT) represents a critical computer vision task that underpins behavioral analysis in intelligent monitoring systems for smart aquaculture applications. However, captured video data often exhibit highly disparate distributions due to variations in fish species characteristics and environmental background complexity. Current fish tracking methodologies frequently rely on suboptimal transfer learning from pretrained models, while existing domain adaptation (DA) techniques inadequately exploit low-level features and multiscale instance-level representations. To address these fundamental limitations, this article introduces an unsupervised DA framework for MFT (DA-MFT). The proposed framework first employs a domain style correction module to enhance feature learning on unlabeled target domain data through integrated low-level feature fusion, thereby facilitating effective target domain-style contextual learning. Subsequently, a novel multiscale cross-head alignment block is developed to capture domain-invariant fish features at the instance level. The framework further incorporates a deep cascade matching strategy to achieve robust and self-contained fish tracking capabilities. To comprehensively evaluate performance across diverse industrial aquaculture scenarios, a specialized domain-adapted MFT dataset was constructed and validated. Experimental results demonstrate that DA-MFT significantly outperforms existing state-of-the-art methods, achieving superior performance with multiple object tracking accuracy scores of 86.6% and higher order tracking accuracy scores of 66.5%.
Keywords:
Computer vision
fish tracking
fishery electronic monitoring
unsupervised domain adaptation
Journal
IF:
9.9
Papers:
8.3K
Citations:
6.0W