Trap Tracker UK Wildlife Detector: YOLO Weights for 31-Class Camera Trap Object Detection

Description:

An open-source object detection model for UK camera trap imagery, released as part of the Trap Tracker platform (traptracker.co.uk). A YOLO-family detector (YOLOv10x / YOLO26x variants) trained to localise 31 classes in a single pass: 28 common UK mammal and bird species, plus utility classes for people, calibration poles, and vehicles. Unlike species-agnostic detectors (e.g. MegaDetector) that return only "animal/person/vehicle," this model returns UK species-level labels directly at the detection stage, so it can be used standalone without a separate downstream classifier. Weights are released in ONNX format (including an FP16 variant) for offline, cross-platform inference.

Training data:

A dataset is used that contains 48,165 labelled animal instances drawn from UK camera trap images collected through Conservation AI and its successor, Trap Tracker, across multiple partner sites over nearly a decade of operational deployment. The dataset was class-stratified and split 80/10/10 into training (38,522 instances), validation (4,805 instances), and test (4,838 instances) sets, with every class — including the smallest, Curlew chick (274) and Calibration Pole (151) — represented in all three splits. Training data spans daylight RGB, infrared night-time monochrome, motion blur, partial occlusion, multi-animal frames, and manufacturer overlays/timestamps.

Raw training images are not released, as a substantial proportion are held under partner-specific data agreements that pre-date this open release.

Classes (31): hedgehog, grey squirrel, red squirrel, roe deer, red deer, red fox, European badger, Person, domestic goat, European rabbit, wood pigeon, common pheasant, domestic sheep, house sparrow, fallow deer, domestic cattle, domestic cat, pine marten, domestic dog, Calibration Pole, northern goshawk, common buzzard, capercaillie (cock), capercaillie (hen), Eurasian curlew, curlew chick, Car (reserved, unpopulated), domestic horse, carrion crow, wild boar, Reeves's muntjac.

Training setup:

Architecture: YOLO-family extra-large variant (YOLOv10x and YOLO26x releases available — performance is comparable between the two); Hardware: 8× NVIDIA RTX A6000 GPUs, global batch size 256; Input resolution: 640×640, mixed-precision (AMP); Framework: Ultralytics/PyTorch; Export format: ONNX (FP32 and FP16)

Performance (validation split, F1-optimal confidence threshold ~0.62):

On the held-out test split (4,838 instances), the model produced confident detections in 4,812/4,820 images across all 31 classes, mean per-species confidence 0.959–0.988, with a false-negative rate of 0.17% concentrated in night-time, distant, or occluded conditions. All metrics are in-distribution (same site/camera pool as training); cross-site generalisation has not yet been benchmarked and is identified as priority future work.

Limitations: Class-stratified (not site- or time-disjoint) hold-out — real-world performance at novel sites/cameras may be lower. Roe deer/red deer confusion is the most common error mode. Not intended as a general UK avifauna detector (bird coverage is a curated subset, not comprehensive).

Intended use: A single-stage, species-level object detector for conservation practitioners, ecologists, and citizen science projects — usable on its own or as a front-end filter ahead of a downstream classifier. Designed to run on a CPU laptop with no ML expertise required. Free for non-commercial conservation, research, teaching, and citizen science use.

Related publication: arXiv:2606.10940, "Democratising Camera Trap AI: An Open-Source Model for Detecting UK Mammals" ("is supplement to")

License: CC BY-NC 4.0

Citation: Please cite arXiv:2606.10940 when using this model in published work. Commercial use requires separate permission (contact Trap Tracker).