A method for automatic honey bees detection and counting from images with high density of bees

This paper presents a design and vision-based techniques for an automated bee counting system. Particularly, the proposed system aims to count bees with high density presences in front of beehive’s entrance. This is a common situation at a bee farm when beekeepers observe honey bee’s appearances to monitor their health. To this end, the proposed system is constructed with a Jetson Nano computer board and a high resolution camera. The bees are automatically and real-time counted from the collected video data. The counting techniques are deployed using recent advantaged deep learning techniques. First, we adapt a YOLO neural network to predict bee’s position on images. However, YOLO is not robust enough in case of occlusions due to a high density of bees’ presence. We then utilize a kernel-based density estimator for each local region. In case a high-density area is detected, we deploy the FAMNET, a neural network recently achieves the best performance for counting objects in high density scenarios. The FAMNET is fine-tuned and optimized parameters for the bee collected data. We measure the performances of the proposed method using a distance between ground-truth counted by beekeepers and the estimated results. The experimental results confirm that it is averagely 10% differences between beekeepers’ counting and the proposed technique. These results show a promising solution to further deploy an IoT system to automatically monitor bee’s health in a bee farm.