Camera traps are an important conservation tool that can provide data on what species are present in a location, the population sizes, and even how species are interacting with each other and the habitat. They are also minimally disruptive to wildlife. However, they generate an enormous number of images or video that can take extensive effort to annotate manually.
Animal Species Captured By Raging River Cameras:
EarthSenseAI is maintaining a small array of camera traps in the Raging River valley in Western Washington. This provides us with direct experience with the challenges of operating these important devices for wildlife conservation. It also provides us with a dataset of video for our machine learning conservation research.
Statistics:
6-10 camera traps in operation at any time
More than 20 species of animals detected
Approximately 6000 animal sightings
Cameras set up to capture 30-60 second videos
First cameras installed in 2019
In spite of the large amount of imagery and video being collected by Washington government studies, native tribes, and non-profit institutions, there are not any specialized machine learning models for recognizing local animal species that can be shared by the various groups utilizing camera traps in the state. Publically available models trained on broad species datasets, such as iNaturalist, do not provide high accuracy for the specific species in WA state. A state specific model has the potential to help multiple organizations pursue machine learning automation in order to obtain substantial, long-term reduction in the enormous amount of human labor in annotating images and video from camera traps.
Coming Soon: Status update on EarthSenseAI neural network model training for detecting and recognizing Washington State species in images and video from camera traps.
Future: Training of a promptable video transformer model, behavior recognition, and integrated video and audio processing.
