Automated nocturnal insect monitoring in real-time using YOLOv4

Details

Supervisors

Bonaventure, Olivier ; Detry, Renaud

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

The sharp decline in insect population during the last decades has highlighted the need for more insect monitoring. Until recently, acquiring data on insect population was done manually and required a significant amount of human workload. To lessen this load, new methods of data collection have been developed using computer vision. Automated systems can recognise insects with the help of AI and monitor the fluctuation of biodiversity in a location. These systems usually consist of a trap, or a drone, to take pictures of the insects. Those pictures are then sent to a central server that performs the recognition. In our thesis, we chose a different approach by performing every aspect of monitoring directly in our trap, i.e. attracting, taking pictures of and recognising insects. With this difference, we had the possibility to make a completely autonomous trap, as it will work offline. We propose a real-time autonomous monitoring trap that uses YOLOv4 as its object detection model with a ResNet18 backbone. The model has been trained and tested on more than 48000 images taken from GBIF. Our experimental results show that detection on the edge is possible in real-time and with an mAP of 85.33% with ResNet18.