AviaryMOT: Aviary Attention-based adaptive multi-object tracking of cranes and storks in wetlands

This study focuses on tracking cranes and storks to aid in wetland ecological protection. Multi-target tracking of these birds presents challenges such as frequent occlusions, sudden appearances, and disappearances. To tackle these issues, we propose a novel multi-target tracking algorithm, AviaryMOT, which utilizes a fusion technique that combines shallow and deep features to enhance tracking accuracy and effectiveness. We construct a dataset, BirdTrack, for cranes and storks tracking. In the detecting stage, we proposed Aviary Attention to effectively capture the features of birds, by integrating the Coordinate Attention into the YOLOv8 framework and applying Soft-NMS to improve detection in occluded scenarios. In the tracking stage, the BYTE data association method effectively utilizes similarities between low-score detection boxes and tracking trajectories, enabling the identification of true objects and filtering out background noise. Experimental results show that our method outperforms the state-of-art models, maintaining stable target trajectories while ensuring high-quality detection.