Robust detection and tracking of laboratory animals in large-scale home-cage video recordings

动物实验在许多研究领域仍然是不可缺少的组成部分。当前的欧盟指令（指令2010/63/EU）基于3Rs原则，旨在尽可能取代、减少和改进动物试验。除了由训练有素的人员定期观察和评估实验动物的健康状况外，持续录制的视频数据为分析实验动物的行为和应激感知提供了丰富客观的信息来源。尽管对视频数据进行直接定量分析几乎是不可能的，而且对研究人员来说也太耗时，但使用最先进的计算机视觉算法对实验动物进行自动定位、自动检测解剖标志和分析动物运动轨迹正变得越来越准确，并允许对社会和个体行为进行详细分析。在视频数据中跟踪实验动物可以非侵入性地进行，允许详细的行为研究，而不会因人为干扰而产生偏见。在FOR开发的家庭笼代表了这样一个非侵入性的环境。然而，所获得的量化质量在很大程度上取决于检测和跟踪算法的鲁棒性，即必须在较长时间内准确地跟踪动物。虽然在项目的前几个阶段已经建立了视频记录和自动存档，但对录制的视频数据的处理和分析是我们将在本项目中解决的剩余挑战。典型的实验持续时间为连续数天的视频记录，每次实验有几tb的视频材料，必须完全自动存档和分析。FOR内拟议的子项目nP20旨在改进和扩展这个家庭笼分析管道的最终部分。它包括(1)一种改进的训练数据注释策略，(2)在顶视图图像中可靠检测和分割实验动物的并行算法，(3)一种用于检测对象时间分配的鲁棒跟踪算法，以及(4)一组基于提取的跟踪数据分析动物（社会）行为的定量描述符。所有开发的方法将被定量验证，并最终用于分析在前两个阶段收集的家庭笼实验视频数据(5)。所有的实现都将作为开源软件公开提供给科学界。

Animal experiments are still an indispensable component in many areas of research. The current EU directive on this (Directive 2010/63/EU) is based on the 3Rs principle and aims to replace, reduce, and refine animal testing wherever possible. In addition to regular observation and assessment of the health status of laboratory animals by trained personnel, continuously recorded video data provide a rich and objective source of information for analyzing the behavior and stress perception of laboratory animals. Although direct quantitative analysis of video data is nearly impossible and too time-consuming for researchers, automatic localization of laboratory animals, automatic detection of anatomical landmarks, and analysis of animal movement trajectories using state-of-the-art computer vision algorithms are becoming increasingly accurate and allow detailed analyses of social and individual behavior. Tracking of laboratory animals in video data can be performed non-invasively, allowing detailed behavioral studies without possible bias caused by human interference. The home-cage developed within FOR represents such a non-invasive environment. However, the quality of the quantifications obtained depends significantly on the robustness of the detection and tracking algorithms, i.e., the animals must be tracked without error over longer periods of time. Although video recording and automated archiving have already been established in the previous phases of the project, the processing and analysis of the recorded video data is a remaining challenge that we will address in this project. Typical experiment durations are in the order of several days of continuous video recording with several terabytes of video material per experiment, which must be archived and analyzed fully automatically. The proposed subproject nP20 within FOR aims to improve and extend the final parts of this home-cage analysis pipeline. It includes (1) an improved strategy for annotating training data, (2) parallelizable algorithms for reliable detection and segmentation of laboratory animals in top-view images, (3) a robust tracking algorithm for temporal assignment of detected objects, and (4) a set of quantitative descriptors for analyzing the (social) behavior of animals based on the extracted tracking data. All developed methods will be quantitatively validated and finally used to analyze video data from home-cage experiments collected in the first two phases of FOR (5). All implementations will be made publicly available to the scientific community as open-source software (6).