The Arribada Initiative is actively working to develop in-field early warning systems utilising thermal technologies. This repository hosts the data, software and hardware designs for both our field research and prototype solutions.
We are currently focused on two species;
- Early warning / detection of elephants using microbolometer sensors & edge machine learning
- Early warning / detection of polar bears using microbolometer sensors & edge machine learning
Arribada's thermal elephant programme, in partnership with the Zoological Society of London, has collected a dataset of Asian elephant photos at ZSL Whipsnade Zoo. This dataset is open source (GPLv3) and has been made available for use by anyone wishing to train their own models using Edge Impulse for the ElephantEdge competition, hosted by Hackster.io and SmartParks. Arribada is also a partner of the OpenCollar initiative and is delighted to support the development opportunity to incorporate an open camera system with inference on the edge.
As part of the WWF Human Wildlife Tech Challenge, Arribada was tasked with developing an early warning thermal system capable of detecting Asian elephants and carnivores, specifically tigers & polar bears. We focused on elephants and polar bears first, partnering with the Zoological Society of London and working with ZSL Whipsnade Zoo to capture thermal photographs of the Asian elephant herd using FLIR Lepton 2.5 and 3.5 microbolometer sensors (80x60 and 160x120 respectively). You can read more about our data collection methodology or the original WWF competition here..
Total number of photographs: 75,830 Total size: 550MB
Adobe bridge has been used to organise data into collections based on various variables:
Angle
- -- Front
- -- Rear
- -- Side
Distance
- -- 0-5m
- -- 5-10m
- -- 10-15m
- -- 15-20m
- -- 20-25m
- -- 25m+
Object
- -- Goat
- -- Human
- -- Human & Elephant
- -- Multiple obstructing elephants
- -- Multiple separate elephants
- -- Single elephant
Sensor
- -- Lepton 2.5
- -- Lepton 3.5
If there was more than one elephant in the image and they were at different distances/angles, the image was sorted into all applicable categories. E.g. one elephant 0-5m from camera at a side angle, and another 10-15m and head on to the camera would be sorted into: 0-5m, 10-15m, head & side collections.
You can use the Edge Impulse command line utility to upload the dataset in order to train a model. To install it, follow the Installation guide. You can then use the Edge Impulse Uploader tool to upload the files in this repository.
For example, you may want to train a model to discern between two categories of image: "elephant" and "non-elephant". In this case, you can upload the corresponding images to Edge Impulse with the following commands:
# Upload all the "elephant" images
edge-impulse-uploader --category split --label elephant human-wildlife-conflict/Elephant/Object/single_elephant/*.png
edge-impulse-uploader --category split --label elephant human-wildlife-conflict/Elephant/Object/multiple_separate_elephants/*.png
edge-impulse-uploader --category split --label elephant human-wildlife-conflict/Elephant/Object/multiple_obstructing_elephants/*.png
edge-impulse-uploader --category split --label elephant human-wildlife-conflict/Elephant/Object/human_and_elephant/*.png
# Upload all the "non-elephant" images
edge-impulse-uploader --category split --label non-elephant human-wildlife-conflict/Elephant/Object/human/*.png
edge-impulse-uploader --category split --label non-elephant human-wildlife-conflict/Elephant/Object/goat/*.png
The --category split
flag ensures the data is split automatically into training and test datasets after the upload.