Multiobject Tracking of Wildlife in Videos Using Few-Shot Learning

Area of Science:

• Ecology
• Computer Science
• Artificial Intelligence

Background:

• Camera trapping and video recording are essential tools in wildlife ecology but face limitations with small sample sizes for species identification.
• Current methods struggle with rare species detection and tracking due to insufficient data, hindering ecological studies.

Purpose of the Study:

• To develop an automated wildlife detection and tracking system using few-shot learning to overcome data limitations in camera trap studies.
• To enhance the performance of wildlife detection and tracking by integrating few-shot object detection with trajectory reconstruction.

Main Methods:

• Proposed a multiobject-tracking approach based on the tracking-by-detection paradigm tailored for wildlife.
• Utilized few-shot object detection with camera trap and video data, incorporating synthetic data augmentation and spatial constraints.
• Introduced a trajectory reconstruction module to improve object association and mitigate detection errors.

Main Results:

• The developed approach created a fully automated pipeline for wildlife detection and tracking from video records.
• Experimental results demonstrated improved detection and tracking performance based on standard evaluation metrics.
• The system effectively addressed missed and false detections, optimizing target identification across frames.

Conclusions:

• The integration of few-shot learning and trajectory reconstruction offers a robust solution for wildlife detection and tracking with limited data.
• This automated pipeline shows significant potential for advancing wildlife behavior and conservation research using camera trap technology.
• The method enhances the utility of camera traps for ecological monitoring and species identification.