Animal re-identification in video through track clustering.

Area of Science:

• Computer Vision
• Animal Behavior Analysis
• Machine Learning

Background:

• Automated animal re-identification from video is crucial for monitoring.
• Multiple Object Tracking (MOT) alone has limitations in accurate re-identification.
• Existing animal datasets are insufficient for training advanced conventional clustering models.

Purpose of the Study:

• To develop an effective automated animal re-identification method for video data.
• To augment Multiple Object Tracking (MOT) with a novel clustering approach.
• To address the limitations of small animal datasets for conventional clustering.

Main Methods:

• Propose a Classification-Based Clustering (CBC) method tailored for animal re-identification.
• Utilize track labels and temporal constraints to train a bespoke model for each video dataset.
• Compare CBC performance against MOT alone and 13 alternative clustering methods.

Main Results:

• The proposed CBC method significantly improves animal re-identification compared to using tracking results alone.
• CBC outperforms 13 alternative clustering methods when applied to tracking data.
• The bespoke model training approach adapts effectively to different video datasets.

Conclusions:

• Classification-Based Clustering (CBC) offers a robust solution for automated animal re-identification in videos.
• The CBC method provides a viable alternative for datasets with limited training data.
• This approach enhances the accuracy and efficiency of animal monitoring systems.