A system for long-term high-resolution 3D tracking of movement kinematics in freely behaving animals

项目摘要 该提案的目的是提供一个创新和易于使用的实验平台，用于测量 并量化用于生物医学研究的哺乳动物模型的自然行为， 包括啮齿类动物和猴子，跨越一系列空间和时间尺度。这将需要发展 一种用于以高得多的时空分辨率跟踪自由行为动物的运动的方法， 更多的运动学细节。为了克服现有技术的局限性， 提出了一种解决方案，该解决方案协同地结合了两种方法-基于标记的运动捕捉和视频， 基于机器学习的方法首先，使用基于标记的动作捕捉，这是3D的黄金标准， 在人体跟踪中，实验对象的头部，躯干和四肢的位置将在3D中被跟踪， 亚毫米精度创新的标记设计、放置策略和后处理管道 将确保在很大的时间范围内对啮齿动物行为进行前所未有的详细描述。到 使系统更加高效、稳健、经济实惠，更适合高通量纵向 研究，运动捕捉实验产生的前所未有的丰富和庞大的3D数据集将 利用深度神经网络来从一组1-6个法线预测姿势和附肢位置， 摄像机为了最好地利用大型训练数据集，卷积神经网络的最新进展 将纳入图像分析网络。总之，这些进步将促进 高分辨率的3D跟踪系统，以各种动物和环境，从而建立了一个便宜， 灵活且易于使用的运动学跟踪方法，可以轻松扩展并被其他实验室采用。 大型地面实况数据集将允许系统进行基准测试，并与最先进的 艺术技术在定量和严格的方式。初步研究结果非常积极， 与当前方法相比，无论是在可跟踪的行为范围方面， 以及它们的测量精度。重要的是，所有新技术都将随时共享 与科学界，从而利用这一单一的赠款的潜力， 研究人员大大提高他们的研究计划的效率，需要严格的 对动物行为的定量描述。

PROJECT SUMMARY The aim of this proposal is to deliver an innovative and easy-to-use experimental platform for measuring and quantifying naturalistic behaviors of mammalian animal models used for biomedical research, including rodents and monkeys, across a range of spatial and temporal scales. This will require developing a method for tracking movements freely behaving animals with far higher spatiotemporal resolution and more kinematic detail than currently possible. To overcome the limitations of current technologies, a new solution is proposed that synergistically combines two methods - marker based motion capture and a video- based machine learning approach. First, using marker-based motion capture, the gold standard for 3D tracking in humans, the position of experimental subjects' head, trunk, and limbs will be tracked in 3D with submillimeter precision. An innovative marker design, placement strategy, and post-processing pipeline will ensure an unprecedentedly detailed description of rodent behavior over a large range of timescales. To make the system more efficient, robust, affordable and better suited for high-throughput longitudinal studies, the unprecedentedly rich and large 3D datasets generated by the motion capture experiments will be leveraged to train a deep neural network to predict pose and appendage positions from a set of 1-6 normal video cameras. To best capitalize on the large training datasets, the latest advances in convolutional neural networks for image analysis will be incorporated. Together, these advances will promote generalization of the high-resolution 3D tracking system to a variety of animals and environments, thus establishing a cheap, flexible, and easy-to use kinematic tracking method that can easily be scaled up and adopted by other labs. The large ground-truth datasets will allow the system to be benchmarked and compared against state-of-the art technologies in quantitative and rigorous ways. Preliminary studies have been very positive and suggest large improvements over current methods both when it comes to the range of behaviors that can be tracked and the precision with which they can be measured. Importantly, all new technology will be readily shared with the scientific community, thereby leveraging from this single grant the potential for numerous investigators to dramatically improve the efficiency of their research programs requiring rigorous quantitative descriptions of animal behavior.