Tracking Small Birds by Detection Candidate Region Filtering and Detection History-aware Association

Tingwei Liu,Yasutomo Kawanishi,Takahiro Komamizu,Ichiro Ide
DOI: https://doi.org/10.48550/arXiv.2405.17323
2024-05-28
Abstract:This paper focuses on tracking birds that appear small in a panoramic video. When the size of the tracked object is small in the image (small object tracking) and move quickly, object detection and association suffers. To address these problems, we propose Adaptive Slicing Aided Hyper Inference (Adaptive SAHI), which reduces the candidate regions to apply detection, and Detection History-aware Similarity Criterion (DHSC), which accurately associates objects in consecutive frames based on the detection history. Experiments on the NUBird2022 dataset verifies the effectiveness of the proposed method by showing improvements in both accuracy and speed.
Computer Vision and Pattern Recognition
What problem does this paper attempt to address?
This paper attempts to solve the problem of tracking small bird - like objects in panoramic videos. Specifically, when the target object (such as a small bird) appears very small in the image and moves rapidly, the existing object detection and association methods will encounter difficulties. To meet these challenges, the authors propose two methods: 1. **Adaptive Slicing Aided Hyper Inference (Adaptive SAHI)**: This method improves efficiency by reducing the number of candidate regions that need to be detected. It samples the candidate regions based on the target positions detected in the previous frame, thereby reducing unnecessary processing. 2. **Detection History - aware Similarity Criterion (DHSC)**: This method more accurately associates target objects in different frames by combining the detection history and the predicted positions. Especially in the case of occlusion (such as when a bird enters a nest), DHSC can utilize the detection history information to improve the accuracy of association. Through experiments on the NUBird2022 dataset, the authors verify the effectiveness of the proposed methods in terms of accuracy and speed. The experimental results show that, compared with the baseline methods, the proposed methods significantly improve the multi - target tracking accuracy (MOTA) and the number of identity switches (IDsw), while reducing false positives (FP) and false negatives (FN), and the processing speed is faster (FPS).