Towards Visual Insect Camera Traps

Thiele Sebastian, Haalck Lars, Struffert Marvin, Scherber Christoph, Risse Benjamin

Abstract

Camera traps have become a standard tool to survey wildlife distribution, abundance and behaviour. Unfortunately, the detection mechanisms are neither sensitive nor selective enough to trigger in case of insect visitations so that current systems can only be used for bigger vertebrates. In this progress report, we present our effort towards a visual insect camera trap. In particular, we discuss why current systems fail, summarise the involved challenges, trained several models on a novel realistic wildlife insect dataset and present the results of our current prototype. Our dedicated deep learning based small object detectors achieve an average precision of 78% while being trained on colour and motion features to identify insects within the field of view of the camera. Finally, we discuss which technical requirements and steps will be necessary to provide a versatile tool for future behavioural, ecological and agricultural studies.