In recent years, the importance of behavioral studies of model organisms such as Drosophila melanogaster has significantly increased in biological research. Recently, a novel monitoring setup for analyzing Drosophila larvae in culture vials was proposed which allows researchers to conduct long-term studies without disturbing the animals' behavioral routine. However, when monitoring larvae in such a setup over several days, dirt accumulates on the vial surface, leading to artifacts in the segmentation process. To overcome this problem and enable researchers to perform experiments involving long-term tracking of the animals, we propose a method for background subtraction which is based on convolutional neural networks (CNNs). Our method produces good results and significantly outperforms other methods. In addition, we show that besides its good performance our compact CNN architecture allows us to apply our method for online-processing on microcomputers in real-time.
展开▼
