UWOT-Net: Underwater Object Tracking by Attention Driven Network in Unconstrained Marine Environments

Understanding and monitoring underwater ecosystems is crucial for ecological conservation, biodiversity studies, and fisheries management. However, tracking marine life, particularly fish, in uncontrolled underwater environments present significant challenges due to appearance variations, motion complexities, and spatial distortions. In this work, we propose an attention drive network, namely UWOT-Net for robust single and multiple underwater object tracking. UWOT-Net integrates image enhancement module with transformer-based tracking architecture. The image enhancement module helps in addressing the underwater degradation. Unlike conventional methods, UWOT-Net utilizes an encoder-decoder transformer framework to enable accurate data association and trajectory prediction across video sequences through a frame-to-frame set prediction mechanism. The transformer framework enables global context understanding and identity preservation without explicit modeling of motion and appearance cues. Experimental evaluations across BrackishMOT, UTB180, and GMOT40 datasets show-case remarkable efficacy of UWOT-Net, attaining state-of-the-art performance in Multiple Object Tracking Accuracy (MOTA) while minimizing ID switches. Notably, our model achieves 100% MOTA score for 8 sequences and surpasses 80% MOTA for 13 sequences, with a remarkable increase of over 100% MOTA for the Brackish dataset. Moreover, UWOT-Net exhibits adaptability by successfully tracking categories beyond fish, indicating its potential for diverse marine and generic underwater object tracking applications.