Ungulate Detection and Species Classification from Camera Trap Images Using RetinaNet and Faster R-CNN.

Area of Science:

• Ecology
• Computer Science
• Wildlife Management

Background:

• Ungulate population density significantly impacts ecosystems and human societies.
• Non-invasive camera trap networks are used for wildlife monitoring but generate vast amounts of data.
• Manual image analysis for animal detection is resource-intensive and time-consuming.

Purpose of the Study:

• To present a novel dataset of wild ungulates from Latvia.
• To develop and evaluate deep learning models for automated ungulate detection in camera trap images.
• To investigate the impact of training optimization and data augmentation on model performance.

Main Methods:

• Collected and curated a new dataset of wild ungulate images from Latvia.
• Implemented and trained two object detection models: RetinaNet and Faster R-CNN.
• Applied data augmentation techniques and analyzed training optimization strategies.

Main Results:

• Demonstrated the effectiveness of RetinaNet and Faster R-CNN for detecting wild ungulates in the collected dataset.
• Quantified the performance improvements achieved through optimized training and data augmentation.
• Validated the models' performance on real-world data from Latvia.

Conclusions:

• Automated detection systems using deep learning significantly improve the efficiency of wildlife monitoring.
• The developed dataset and models provide a valuable resource for ungulate population studies.
• Further research can build upon these methods for enhanced wildlife management and conservation efforts.