Autonomous UAV Navigation for Individual Wildlife Re-Identification

Claire Sun, Tanya Berger-Wolf, Jenna Kline · 2026

Read Autonomous UAV Navigation for Individual Wildlife Re-Identification by Claire Sun, Tanya Berger-Wolf, Jenna Kline (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-31772.txt.

Introduction

Sun, Berger-Wolf and Kline design a drone that flies itself to photograph individual wild animals for re-identification. Recognizing one animal from its coat pattern needs a clear side view at good resolution. Ecologists and citizen scientists now gather such images by hand, which is slow and uneven in quality.

The system pairs YOLOv11 detection with a pose classifier built on DINOv2. It finds an animal, turns to expose the lateral flank, then approaches until the subject fills enough of the frame. A case study on zebras uses footage collected in Kenya. The authors argue the same task-aware approach fits giraffes, tigers and elephants, whose surface patterns also identify individuals.

Catalog record

CREATORClaire Sun, Tanya Berger-Wolf, Jenna Kline
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.31772

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Autonomous UAV Navigation for Individual Wildlife Re-Identification Claire Sun, Tanya Berger-Wolf, Jenna Kline Abstract Reliable individual re-identification (re-ID) of wildlife is essential for population monitoring, behavioral tracking, and conservation policy evaluation, yet large-scale data collection remains labor-intensive, relying on manual efforts by ecologists or citizen scientists. We propose an autonomous drone navigation system that actively optimizes image capture for downstream re-ID, moving beyond passive aerial sensing. The system combines YOLOv11 object detection with a DINOv2-based pose classifier to guide real-time flight decisions: detecting animals, orienting to expose the lateral flank (the surface of interest for pattern-based re-ID), and approaching until the subject meets a minimum bounding-box threshold. Unlike prior drone systems that optimize for group-level behavioral video, ours targets the specific image-quality requirements of individual-identification models. We demonstrate feasibility through a case study on zebra using footage collected in Kenya, and show the approach generalizes to other species with diagnostic surface patterns, including giraffes, tigers, and elephants. Our work establishes a framework for task-aware embodied AI for ecological data collection, in which downstream re-ID requirements drive real-time perception and control. 1 Introduction Accurate population counts and longitudinal tracking of individual animals are foundational to conservation science. They underpin management decisions for endangered species and enable rigorous evaluation of protection policies. The Great Grevy’s Rally, a citizen-science census of Grevy’s zebra ( Equus grevyi , fewer than 3,000 remaining in the wild), illustrates both the value and the cost of this work: mobilizing hundreds of volunteers for a single multi-day survey 7 . Individual re-identification (re-ID) of wildlife from natural markings, pioneered by systems such as HotSpotter 6 and now extended by deep learning approaches 28 ; 4 ; 20 ; 24 , has significantly reduced the need for invasive tagging. Platforms like Wildbook aggregate citizen-science images to track individuals across time and geography 3 . However, the data collection bottleneck remains: images must be manually gathered, and their quality is highly variable. An image is only useful for re-ID if it captures the correct surface-of-interest (e.g., the lateral flank for zebras) at sufficient resolution. These conditions are rarely guaranteed in opportunistic photography. Unmanned aerial vehicles (UAVs) have transformed wildlife monitoring by enabling rapid, large-area surveys with reduced disturbance 5 . Recent work has moved from manual to autonomous UAV operation for behavioral studies 17 ; 22 ; 11 demonstrating that integrating ecological knowledge into navigation policies improves data quality 12 . However, these systems optimize for group-level behavioral video by maintaining altitude and herd coverage, rather than for individual identification, which demands fundamentally different image properties. This paper addresses that gap with three contributions. (1) A task-driven navigation protocol (detect, orient, approach, and capture, or DOAC) that couples pose classification and flight control to maximize re-ID-quality image yield. (2) A DINOv2-based lateral pose classifier for zebra flank visibility, deployable on edge hardware. (3) Feasibility validation on real Kenya UAV footage, with a path to generalization across species with diagnostic surface patterns. Figure 1 : Navigation pipeline for individual zebra identification employing the detect-orient-approach-capture (DOAC) approach. The zebra is detected using YOLOv11, then its pose is classified using DINOv2. If a lateral view is detected, the drone moves forward until the bounding box (bbox) threshold is reached. If there is not a lateral view, such as the front-right view shown, the drone yaws left or right until a lateral view is captured, or the timeout condition is reached. 2 Related Work Animal re-identification. Re-identification of individual animals using computer vision has been demonstrated for zebras, tigers, elephants, giraffes, and many other species with distinctive coat patterns or morphologies 28 . Wildbook and ElephantBook provide platforms for managing re-ID at scale 3 ; 19 . Recent work has pushed toward species-agnostic re-ID. Čermák et al. introduce WildlifeDatasets and MegaDescriptor, a foundation model for multi-species re-ID 4 , while Li et al. propose a high-frequency transformer for unified multi-species identification 20 . Nepovinnykh et al. combine learnable local features with aggregation for robust pattern matching across small datasets 24 . A consistent theme across these methods is their sensitivity to image quality, as accurate identification requires a well-defined pose and sufficient bounding-box resolution. Drone-based wildlife monitoring. UAVs have become standard tools for detection-focused wildlife surveys 5 . BuckTales introduces a multi-UAV re-ID and tracking dataset for wild blackbuck antelope 23 The In Situ Kenyan Animal Behavior Recognition (KABR) dataset provides a large collection of video clips of labeled zebra and giraffe behaviors 10 . Andrew et al. equip a UAV with onboard deep inference for biometric identification of individual Friesian cattle using a detector, an exploratory navigation network, and an identification network 1 . Our work differs from this foundational work in targeting flank-based morphologically distinctive re-ID, explicitly formalizing pose as a navigation objective, and using a modern DINOv2-based classifier. Rolland et al. formalize surface-of-interest (SoI) coverage for drone swarms 27 , and Kline et al. identify SoI as a critical planning variable for AI-driven aerial ecology missions 11 . No existing system operationalizes SoI at the individual animal level in real time. Our drone navigation protocol does exactly this by actively seeking the lateral-flank SoI for each detected individual, closing the loop between mission-planning concepts and autonomous execution. Autonomous UAV navigation for ecology. WildBridge 22 demonstrates accurate animal localization from commercial UAVs but treats image capture as passive. Thus, optimizing the viewpoint for identification quality remains unaddressed. WildWing 17 and the imageomics framework of Kline et al. 12 demonstrate autonomous navigation policies derived from expert pilot telemetry, but optimized for group behavioral video, not individual identification missions. Grushchak et al. present decentralized multi-drone herd coverage 8 . These systems treat data quality as a function of altitude, distance, and angle to the group. In contrast, our system defines quality at the individual level through pose and resolution constraints, which is a necessary shift for re-ID. 3 Method 3.1 Problem Formulation An image is re-ID suitable for a given individual if two conditions hold. One, the surface of interest is visible. For zebras, the lateral flank bears the unique stripe pattern. Two, the bounding box occupies sufficient image area for pattern matching. We use a minimum threshold of 500×500 pixels, following standard re-ID practice 11 . Our navigation protocol is designed to maximize the yield of images satisfying both conditions. The choice of lateral flank over top-down view is deliberate. While nadir (top-down) imagery is standard for population counting, individual re-ID relies on lateral flank patterns. This is the view from which expert field researchers identify animals, and from which the majority of curated re-ID training data (camera traps, ground-level photography) has been collected 28 ; 3 . Targeting the lateral SoI allows our system to directly leverage existing annotated datasets and pretrained re-ID models, rather than requiring new data-collection paradigms. However, if future models are developed to conduct individual identification from the nadir view, this framework can be easily adapted to accommodate. 3.2 Navigation Pipeline The system operates in four sequential stages, executed in a real-time loop on onboard compute, illustrated in Fig. 1 . 1. Detect. A YOLOv11 model 9 fine-tuned on aerial zebra imagery processes each video frame. Detected bounding boxes are ranked by confidence. The largest (nearest) detection is

Read the complete text