Abstract: Traditional bee monitoring mostly relies on labourer and experience, and the level of information is low, so automatic monitoring of bees has been concerned widely. Recently, object detection based on deep learning has developed rapidly and achieved good results in many fields. SSD model is a target detection model based on convolutional neural network, which has the advantages of speed and accuracy. The light of the beehive gate area is changeable, and the environment are complex. Bee colony also has complex situations such as mutual occlusion and shadow.  In this paper, SSD model was used to detect honeybees in the hive gate  and to make data statistics analysis. The results showed that the precision of the method selected was 96.34%, 92.52% and 88.06% in the case of small, general and large bees’ number, which was 11%, 19% and 25% higher than that of the traditional method, respectively. Furthermore,  SSD model had strong adaptability to different environment such as illumination, weather, shooting distance and could detect and deal with complex conditions such as bee shadow, blur, occlusion, etc. The detection results provided strong support for monitoring outside the hive, and also laid a foundation for statistics of in-and-out number based on bee tracking.

Key words: hive gate area, bee detection, deep learning, SSD

摘要： 传统蜜蜂监测多依靠人力和经验，信息化水平低，蜂群自动化监测得到广泛关注。近年来，基于深度学习的目标检测发展迅速，并在多领域取得很好的应用效果。SSD模型是一种基于卷积神经网络的目标检测模型，具有快速和准确率高的优势。蜂巢口光照多变、环境复杂，蜂群本身也具有互相遮挡和阴影等复杂情况。采用SSD模型对巢门区蜜蜂检测和数据统计，结果表明，提出的方法在少量、一般和较多蜜蜂数量情况下准确率分别达到96.34%、92.52%和88.06%，比传统方法分别提高11%、19%和25%，且对光照、天气、拍摄距离等环境有很强的适应性，能检测处理蜜蜂阴影、虚化、遮挡等复杂状况。研究结果为蜂群巢外监测提供有力支持，也为基于蜜蜂跟踪的进出量统计奠定了基础。

关键词: 巢门区域, 蜜蜂数量, 深度学习, SSD