Aplicando modelos de visão computacional para detectar em tempo real o fluxo de entrada de pólem em colméias de abelhas melíferas

Abstract

Pollination is a critical process for the reproduction of plants and the formation of seeds and fruits. Bees are widely recognized as the main agricultural pollinators, and according to literature, the pollen foraging task into hives is directly associated with their health. Hive and colony monitoring is traditionally carried out through manual and invasive checks, which stress the bees and can be ineffective, as it depends on the beekeeper’s experience. On the other hand, the advent of the recent precision beekeeping paradigm has enabled remote and non-invasive monitoring of hives. Despite the vast literature on machine learning models to detect pollen caps on bees, there are restrictions in the data collection that consequently results in a dataset with low variability. From videos recorded at honeybee hive entrances, in this dissertation, we apply six state-of-the-art computer vision models to detect the pollen foraging task: Deformable DETR, Faster R-CNN, Cascade R-CNN, YOLOv7, YOLOX, and CenterNet. To ensure that the dataset was representative of this task and as agnostic as possible to the capture environment, we generated a set of images, collected from over 15 different sources, including variation in camera positioning, lighting, and contrast. Additionally, three basic image characteristics based on model predictions were studied: contrast, texture, and coloration. The results showed that YOLOv7 and Deformable DETR were statistically equivalent in the experiments, both obtaining metrics above 77% on the most restrictive and challenging AP75 metric, in terms of details, among those analyzed in the work. With regard to the FPS (frames/s) metric, important for real-time detection, YOLOv7 provided a value four times higher than obtained by the Deformable DETR. The CenterNet model presented the best metrics for real-time applications, with good predictive performance at a low computational cost, obtaining values of approximately 94% on the AP50 metric with about 20% of the computational cost of the YOLOv7 model.