自然状況下インタラクション解析に向けた動物集団の映像に対する姿勢追跡

村重 哲史


The social behavior analysis of animals is needed in various research fields. To observe animal behavior under natural conditions, computer vision-based tracking methods are an effective approach to avoid distraction to animals. Although methods for single-animal pose estimation have been proposed, there are few methods for multi-animal pose tracking that is indispensable to analyze social behaviors. We proposed a method for multi-animal pose tracking using transfer learning of multi-person pose estimation method to images of multi-animal videos with a tracking method. In an experiment, the proposed method was applied to two datasets, one with annotations of individuals' ID and the other without. Our experimental results suggest that our method can analyze videos that single-animal pose estimation methods cannot be applied to.