Open-source AI model for Sub-Saharan Africa camera trap species detection
The TrapTracker Sub-Saharan Africa Mammals Model is an object detection model designed to support camera trap image analysis across Sub-Saharan African wildlife monitoring projects.
The model has been trained to detect a wide range of African mammal species commonly captured in camera trap studies, alongside people and vehicles. It is intended to help conservation organisations, researchers, students, and field teams process large image datasets more efficiently while keeping analysis workflows local, transparent, and accessible.
The model is released for non-commercial conservation, research, education, and citizen science use.
Model Overview
The model is based on the YOLO26x object detection architecture and was trained using camera trap imagery representing a broad range of Sub-Saharan African mammal species.
It is designed to detect and localise animals within camera trap images, including images captured under challenging field conditions such as low light, infrared night imagery, partial occlusion, distant subjects, multiple animals in a frame, and varied habitat backgrounds.
The model is intended as a practical tool for accelerating image review, reducing manual sorting time, and supporting conservation monitoring workflows.
What the model detects
The model includes the following classes:
- Southern African cheetah
- Car
- Blue wildebeest
- Plains zebra
- Baboon
- Giraffe
- African elephant
- Lion
- Person
- Rhinoceros
- African buffalo
- Common eland
- Spotted hyena
- Impala
- Black-backed jackal
- Roan antelope
- European rabbit
- Common warthog
- Chimpanzee
- Giant pangolin
- Crested porcupine
- Aardvark
- Hippopotamus
- Gemsbok Common ostrich
- Hartebeest
- Waterbuck
- Gorilla
- Bongo
- Kob
- Helmeted guineafowl
- Common duiker
- Genet
- Giant pouched rat
- Tantalus monkey
- Leopard
Training Data
The model was trained on labelled instances drawn from Sub Saharan Africa camera trap images collected through conservation deployments over nearly a decade.
The dataset includes a wide range of practical camera trap conditions, including:
- daylight RGB images
- infrared night-time images
- motion blur
- partial occlusion
- multi-animal frames
- juveniles and adults
- manufacturer overlays and timestamps
- varied Sub Saharan Africa habitats and deployment settings
Training setup
The model was trained using:
- Architecture: YOLO26x
- Input resolution: 640 × 640
- Training hardware: 8 × NVIDIA RTX A6000 GPUs
- Training schedule: 120 epochs
- Global batch size: 256
- Framework: Ultralytics
- Export format: ONNX
- Training time: approximately 9 hours 33 minutes
The model was trained using an 8-GPU setup and a large global batch size to support efficient training across the full dataset.
Performance
On the validation split, the final model achieved:
| Metric | Value |
|---|---|
| Precision | 0.971 |
| Recall | 0.964 |
| mAP@0.5 | 0.982 |
| mAP@0.5:0.95 | 0.907 |
| Best F1 score | 0.97 |
| Best F1 confidence | 0.471 |
The training curves show stable convergence across the 120 epochs. Precision, recall, and mAP increased rapidly during early training and then stabilised, while training and validation losses continued to decline.
The F1-confidence curve suggests that a confidence threshold of approximately 0.47 provides the best overall balance between precision and recall on the validation data. For practical camera trap workflows, users may wish to adjust this threshold depending on whether they prefer fewer false positives or fewer missed detections.
On the held-out validation split, the model achieved strong overall performance, with precision of 0.971, recall of 0.964, mAP@0.5 of 0.982, and mAP@0.5:0.95 of 0.907. The F1-confidence curve indicates a best overall F1 score of 0.97 at a confidence threshold of 0.471.
These results suggest strong in-distribution performance across the Sub-Saharan Africa camera trap detection task, while still requiring local validation when applied to new sites, camera types, or species distributions.
Important Limitations
Although the model performs strongly on the validation data, users should treat the results as a practical baseline rather than a guarantee of performance in every setting.
Performance may vary when applied to:
- new reserves or field sites
- different camera trap models
- unusual lighting or weather conditions
- species not included in the class list
- rare behaviours or unusual poses
- heavily occluded animals
- very small or distant animals
- images from regions outside the training distribution
The model should be tested on a small sample of local data before being used at scale in a new project.
Intended Use
This model is intended for non-commercial use in:
- Sub-Saharan African camera trap image analysis
- wildlife monitoring
- biodiversity surveys
- ecological research
- conservation project workflows
- university research and teaching
- citizen science projects
- offline or local image processing
- pre-filtering images before manual review
The model is particularly useful where users need to process large numbers of images and identify likely species detections before manual verification.
Not Intended For
The model is not intended for:
- commercial resale
- paid inference services
- use as a legally authoritative identification system
- replacing expert ecological judgement
- surveillance or monitoring of people
- use outside the model’s species coverage without validation
- making conservation decisions without appropriate human review
Where outputs are used in publications, reports, conservation management, or policy-facing work, results should be manually checked and validated.
Why we are releasing it
Camera trap AI has become increasingly important for biodiversity monitoring, but access to high-performing species recognition tools is often limited by commercial platforms, proprietary weights, pay-per-image models, or cloud-based inference.
This release is intended as a practical counterweight to that trend.
The aim is simple: a conservation organisation with limited funding, no machine-learning team, and no GPU should still be able to process camera trap images locally and obtain useful species-level detections.
The model is free for non-commercial use, runs offline, and is designed for ecologists and conservation practitioners rather than machine-learning specialists.
How to cite
If you use the model in research, reports, publications, or conservation outputs, please cite:
Paper to follow
A formal citation will be added here once the paper is available online.
Licence
The model weights are released for non-commercial use.
You may use the model for conservation, research, teaching, ecological monitoring, citizen science, and non-commercial biodiversity work.
You may not resell the model, use it to provide a paid commercial inference service, or incorporate it into a commercial product without prior permission.
For commercial licensing enquiries, please contact:
Download
By downloading the model, you agree to use it only for non-commercial purposes and to acknowledge the model appropriately in any resulting publications, reports, or public outputs.
Updates
The Trap Tracker Sub Saharan Africa Mammals Model will continue to be developed as new data become available and as the underlying model architecture improves. Future releases will be made available periodically, with clear versioning so that users can track changes, improvements, and compatibility across model updates.