WILDLABS logoWildlabs.net : The conservation technology network

WILDLABS logoWildlabs.net : The conservation technology network

article / 29 August 2025

New Paper: Drones and Computer Vision for Collective Animal Behavior Studies

From detection to behavior, we outline how to run low-impact, reproducible wildlife studies using drones and edge AI.

Groups

We’ve just published a new paper in Methods in Ecology and Evolution on using drones and computer vision to study collective animal behavior. Watch our short video abstract here: Video abstract ▶︎

If you work at the intersection of conservation and technology—or you’re curious how edge AI can make field studies more scalable and less disruptive—this post is for you.

TL;DR

Edge AI on drones and field devices is moving wildlife research from “collect everything and analyze later” to “analyze in place, record what matters, and fly smarter.” Our paper maps the current toolkit (detection → tracking → behavior), shows why mission design and edge compute must be co-designed, and offers practical guidelines for standardized, low-impact data collection.

Key takeaways for edge computing & edge AI
  • Plan for the model, not just the flight. Choose altitude, speed, and viewing angle (often oblique) to maximize downstream tracking and behavior inference, not just detection.
  • Do inference at the edge. Onboard GPUs or nearby field computers can run lightweight models to triage events in real time—record more when behavior is likely, less when it’s not—to save bandwidth, storage, and battery.
  • Temporal models are the next frontier. Video-native architectures (for pose and behavior) help move beyond single-frame detection toward behavior budgets and social interaction mapping in the field.
  • Semi-autonomous flight improves data quality. Consistent, parameterized missions (altitude/speed/safety buffer) standardize datasets and reduce disturbance versus purely manual piloting.
  • Edge optimization matters. Techniques like pruning, quantization, and distillation enable real-time inference on resource-constrained hardware without a satellite link.
  • Measure welfare alongside accuracy. Build safety zones and disturbance metrics into your mission specs; log parameters and context (habitat, weather, group composition) for reproducibility.
  • Think multi-view and multi-modal. Multi-drone (or drone + camera trap + bioacoustics + GPS) meshes reduce occlusion and capture rare events; practical adoption needs accessible swarm tools and simple field UIs.
  • Design for real field conditions. Assume intermittent power/connectivity; favor rugged gear, simple workflows, and on-device fail-safes over cloud dependence.

What’s inside the paper

  • A clear pipeline from detection → tracking → identification → behavior → ecological context.
  • A survey of methods, datasets, and mission designs, with notes on where results vary by species/habitat.
  • A checklist for planning edge-AI-enabled drone missions that are repeatable, low-impact, and analysis-ready.

Read the paper

Full paper in Methods in Ecology and Evolution — start with the figures and the guidelines checklist if you’re skimming.

Questions, datasets, or collaborations? We’d love to hear from you.

Add the first post in this thread.

Want to share your own conservation tech experiences and expertise with our growing global community? Login or register to start posting!