Multiple subject tracking in long-term, behavioral studies of rodents

High-resolution, automated or semi-automated behavioral phenotyping of rodents is increasingly used in neuroscience to reveal behavioral characteristics, that are unreachable with standard manual methods. In fact, these computational approaches have become so remarkable in recent years, it is possible now to extract not only basic ethograms of animals in an observable environment, but also to dissect individual components of particular behaviors. Currently, most computational phenotyping is based on image/video processing. The automation of such behavioral analysis promises consistent, high-throughput, unbiased results, and in general sense that can be true. Nevertheless, when looking more closly at the methods’ workflow and its implementation, some obstacles inevitably emerge. In this article, the issue of identity mismatches in multi-animal, behavioral studies will be discussed, which is especially significant for long-term (multi-hour, multi-day) format. It will briefly explain how in image-based, machine learning approaches, the identity of a subject can be maintained and why algorithms struggle to do so in some situations. In response to those limitations, additional identity-tracking solutions will be presented (ArUco tags, RFiD tags, and color-based recognition) that may significantly improve the overall process of behavioral screening that relies on available open-source tools for pose estimation, such as DeepLabCut or SLEAP.ai.