Advanced wildlife camera trapping using embedded AI machine vision

Conservation of biodiversity is important as a part of safeguarding earth's ecosystems. Collection of biodiversity data is usually done with wildlife cameras equipped with simple sensors based on movement or body heat at remote areas and limited resources regarding power and image storage. A downside of these systems is that they generate many false positive thereby overloading the image storage capacity. In this study we examined the potential of embedded artificial intelligence models to replace traditional camera sensors. To do this, different computer vision models were tested after conducting various experiments to optimize performance. The AIR sensor, which served as a baseline for wildlife cameras, resulted in the worst performance and a lot of false positives. A simple motion sensor and two different image classification models were tested on a data-set, which was collected by a video camera inside a nestbox. These achieved higher performance, but still contained false positives. A final trigger method was designed where an image classification model was run on images that were determined to contain significant amounts of movement. This resulted in a model that could identify nearly all true positives, while reducing false positives to nearly 0 at the cost of a higher power consumption.