The Be-Hive Project—Counting Bee Traffic Based on Deep Learning and Pose Estimation

Abstract

Beekeeping is an important practice for ensuring the abundance of pollinators and for honey production. Traditionally, beekeepers inspect hives regularly to monitor their bees’ populations, but this method is invasive and can cause stress to the bees. It is also impractical, for beekeepers having hundreds or thousands of beehives. In recent decades, various attempts have been made to automate the monitoring of bee colonies using emerging technologies. These technologies include sensors that collect micro-climate parameters, photos, video and audio from inside the hives and the nearby environment, which are then analyzed using automatic or manual methods. The beehive project aims a range of sensing technologies (image, sound, temperature, humidity, weight), together with state-of-the-art computer vision technologies and remote-sensing imagery to create a smart beehive system and monitor beehive on real-time. In this paper, we present the preliminary results of the BE-HIVE, a smart beehive monitoring system. We present the monitoring system developed and the deep learning algorithm used to count bee traffic using the image from the camera placed at the entrance of the hive. For bee traffic estimation, we employ a counting algorithm that predicts the pose of individual bees and tracks them in subsequent frames. To reduce the annotation overhead of the key-points for pose estimation, we generate synthetic data to train our algorithm. The results show that the key-point detection model achieves an Intersection Over Union (IOU) of 86% when trained only on synthetic data and a traffic count mean absolute error of 5.7. These results indicate that the proposed approach can be used to monitor the bee activity remotely, increasing convenience and productivity.