The Use of Video to Detect and Measure Pollen on Bees Entering a Hive

Yang, Cheng (Robert)
Collins, John
Beckerleg, Mark
Lie, Tek Tjing
Item type
Degree name
Doctor of Philosophy
Journal Title
Journal ISSN
Volume Title
Auckland University of Technology

This research will measure the pollen being brought into the beehive in the bees' pollen sacs. It will process 2-D video which is obtained from in front of the entrance to the beehive and automatically count the number of full pollen sacs which bees bring back. This can help bee keepers to check forage inside beehive without opening the hive manually. The technology used in this research relates to object detection using computer vision. Many papers in the field of object detection try to detect and track different objects, such as human beings, vehicles and animals. However, very little research has focused on the flight of bees, and few has involved identifying the pollen sacs. Difficulty arises from requiring high resolution video and high speed processing because bees can fly at high speed.

The first step of the procedure for this research is to detect and track bees. The bee detection method is to combine foreground subtraction and colour thresholding. The tracking approach mainly uses the Kalman filter. After that, the Hough transform is to solve the problem of tracking bees in the occlusion situation.

The second step is the pollen detection and measurement. There are two methods to complete this mission in this research. The first method is to use image processing and statistics to analyse individual bee images. The bee’s body is analysed to identify the main body (head and abdomen) and other parts (wings, legs and pollen sacs). Then the pollen sacs can be detected using the features of colour, size, blob orientation, blob ellipticity, position and blob extent. The thresholds to distinguish pollen and non-pollen blobs are estimated using the receiver operating characteristic (ROC). The amount of pollen is estimated by counting the bees with full pollen sacs. The second method is to use deep learning method. 1000 individual bee images with pollen sacs labels are collect for training a deep neural network. After the training, a network can be used to detect pollen on individual bee images automatically. Then the individual bee images can be identified as pollen or non-pollen images. This identification is combined with bee tracking on bee monitoring video to count the number of pollen carrying bees.

This thesis explains the detail of the theory of the methods used in this research. In addition, it reports the results of applying the methods in practice. The experimental results indicate the tracking of single bees is over 99% accurate. More than 80% of bees in a merged situation are tracked successfully. In pollen detection and measurement test, the difference (or error) between measured number and actual number of pollen sacs is about 7%, if the deep learning method is applied. This is a significant improvement from the image processing and statistics model, which produces 33% error.

Pollen detection , Bee detection , Computer vision , Image processing , Deep learning
Publisher's version
Rights statement