Assessing Rotation-Invariant Feature Classification for Automated Wildebeest Population Counts

Area of Science:

• Ecology
• Computer Science
• Wildlife Management

Background:

• Accurate animal population counts are crucial for effective wildlife conservation and adaptive management.
• Manual counting of animals from aerial imagery is time-consuming and prone to human error.
• Advancements in computer vision offer potential solutions for automated animal detection and enumeration.

Purpose of the Study:

• To implement and evaluate a computer vision algorithm for detecting and counting wildebeest from aerial images.
• To compare the algorithm's accuracy against human enumerators for population counts.
• To identify areas for future research to improve automated animal counting methods.

Main Methods:

• Development and implementation of a rotation-invariant object descriptor-based algorithm.
• Application of the algorithm to aerial images from the Serengeti National Park wildebeest count (2009).
• Comparison of algorithm's per-image and total count accuracy against two independent human counts.

Main Results:

• The algorithm's per-image error rates were comparable to, though slightly higher than, human counts.
• The algorithm achieved higher accuracy than both human counts for the total wildebeest population estimate.
• Findings suggest systematic over- or under-counting tendencies in manual enumeration.

Conclusions:

• Computer vision algorithms show significant potential for automating aerial wildlife population counts.
• The developed algorithm is a promising step towards accurate and on-demand animal enumeration.
• Further research focusing on algorithm refinement and bespoke image collection protocols is recommended for full automation.