Software Beehaviour


  • Beehaviour is a software solution for detecting complex behavioural pattern pertaining to honey bee cell openings.

  • The goal is to identify worker bees which are capable of detecting varroa parasites through the cell cover in order to genetically analyze them. The corresponding cell openings are associated with a distinctive behavioural pattern, most notably oscillating head movements.

  • With the help of the software, we were able to identify significantly more cells in less than half the time compared to manual analysis.

  • The project was realized in collaboration with Prof. Dr. Kaspar Bienefeld at Länderinstitut für Bienenkunde, Hohen Neuendorf (LiB) and funded by the FP7 EU project “SmartBees”

Video demonstrations

Outline of the video processing module

  • The user can load a new bee hive video, which then gets processed by tracking the visible tagged bees and computing the relevant statistics for the subsequent machine learning analysis.

  • In the bee hive recording (left windows in the video), the green circles denote the relevant varroa infested cells. All tracked bees are marked by the colored borders around their tags.

  • The window to the right shows the feature information extracted from the individual bee movements. The green circles again denote the infested cells.

Detailed view of video processing

  • Here, the video processing is shown at a higher detail level.

  • The trajectories of the movements are being recorded in real time as feature information in the window to the right. The shaded green areas at or inside the circles denote head movement activity, which could be indicative of cell opening behaviour.

Outline of the evaluation module

  • Once the video processing of the video is completed, the user can evaluate the results in the evaluation module.

  • For each cell (highlighted in the right window), the user can inspect the timeline of its activity in the recorded video just by moving the mouse over the timeline in the lower left.

  • The detection of relevant behavioural movements by the machine learning algorithm is shown as the pinkn burst in the timeline at the lower left.

Detailed view of the evaluation

  • Here, a detected cell opening event by a worker bee is shown in greater detail. One can spot the relevant behavioural pattern movement of the bee at the cell with its characteristic head wiggling.

  • Once the bee has left the cell cover (at 0:16 in the video), one can clearly see the hole in the cell cover.

  • The relevant pattern was correctly detected by the software, as can be seen by the pink burst in the timeline.


    • Why semi-automatic instead of automatic? Compared to all the bee activity in the honey comb over the course of the video recording the relevant behavioural pattern is an extremely rare event. Here, the big challenge is to detect all the relevant rare events without gathering too many false positives. A further challenge was the limited number of training data, which consisted of cell openings in one bee hive recording with 50 prepared cells. For the purpose of maxmizing the use in practice, we therefore decided to design the software as a support platform for the human expert. Here, the expert can inspect the possible relevant events, which are only a tiny fraction compared ith the complete video and decide if the events pertains to a real cell opening event. This way, the evaluation time can be significantly reduced. Still, next to the reduction in time the algorithm also detected more cell opening events than the human expert, which analyzed the video completely by hand.

    • Why not directly detect the holes in the cell cover instead of analyzing bee movements? The problem with the detection of holes in the cell cover is that due to the large number of bees in the comb, the cell cover is not clearly visible most of the time. Additionally, due to its non-smooth surface, even small variations in lighting conditions and shading effects render this approach unsuitable in practice.

    • Could the software be used for similar applications? Absolutely! Beehaviour can be  adapted to many other detection applications in the biological sciences and beyond. The straighforward integration of human expert knowledge into the learning process allows for efficient and robust learning of complex pattern detection even in applications, where only a moderate amount of training data is available. If you would like to have more information or be interested in discussing a similar application scenario please contact us, we are very happy to help:

Further information