Proceedings of the Annual Meeting of the Cognitive Science Society

Video recording is a widely used tool for studying animal behavior, especially in fields such as primatology. Primatologists rely on video data to analyze and research topics such as social grooming to uncover subtle mechanisms behind complex social behavior and structures. Insights into these social behaviors may provide us with a better understanding of our closest living relatives, but also new theories and insights into our own behavior. However, analyzing this type of data using manual annotation is currently a time-consuming task. Here we present an end-to-end pipeline to track chimpanzee (Pan troglodytes) poses using DeepLabCut (DLC) which then serves as input to a support vector machine. This classifier was trained to detect role transitions within grooming interactions. We replicate a recent study showing that DLC has usability value for chimpanzee data collected in natural environments. Our combined method of tracking and classification is remarkably successful in detecting the presence of grooming, indicating the directionality and a change in turn with an accuracy above 86% on unseen videos. We can identify particular pose features used in the classification of grooming, which will contribute to the exploration of turn-taking dynamics on a scale that would otherwise be difficult to attain with traditional methods. Finally, our pipeline can in principle be applied to recognize a variety of other socially interactive behaviors that are largely recognizable by (joint) postural states.