RI: Small: Clustering, Classification and Alignment of Time Series for Human Sensing

时间序列分析是计算机视觉、信号处理、计算机图形学、机器学习和社会科学等领域研究的核心。 该项目需要对无监督模式发现和时间序列的时间对齐算法及其在人体运动建模中的应用做出原创性贡献。具体而言，PI提出了三种新方法：（1）区分时间聚类，将一组时间序列分解为属于k个时间聚类之一的片段，（2）用于发现两组时间序列之间最具区分性片段的方法，以及（3）用于时间对齐多模态时间序列的无监督算法。PI提出了一个能量最小化框架来解决这三个问题。 这个框架应该为研究人员提供大量现有的时间序列技术的透彻理解，它可以作为一个工具，处理时间序列中出现的其他问题，使计算机能够理解人类行为有可能彻底改变许多领域，造福社会，如临床诊断，人机交互，和社会机器人。从感官数据中模拟人类行为和事件的时间序列的进步对于能够识别和表征人类行为的系统的成功至关重要。然而，大多数现有的算法在本质上是有监督的。监督学习通常需要大量的人工注释，这通常是劳动密集型的，并且难以在编码器之间标准化。在该提案中，PI探索了使用无监督学习技术来对齐和发现使用加速度计，视频或运动捕获技术捕获的人体运动时间序列中的模式。PI将展示所提出的算法如何在几个人类感知任务中优于最先进的技术，例如人类运动的时间对齐，视频中人类活动的时间聚类，学习运动原语以及人类行为的联合分割和分类。 在教育方面，PI将继续为卡内基科学中心提供支持，以展示人类传感技术，并将大量本科生纳入他的研究。研究源代码将提供给科学界。

Time series analysis is central to the study of computer vision, signal processing, computer graphics, machine learning, and social sciences, among other fields. This project entails original contributions towards algorithms for unsupervised pattern discovery and temporal alignment of time series, and its applications to model human motion. In particular, the PI proposes three new methods: (1) a discriminative temporal clustering that factorizes a set of time series into segments belonging to one of k temporal clusters, (2) a method for discovering the set of most discriminative segments between two sets of time series, and (3) an unsupervised algorithm for temporally aligning multi-modal time series. The PI proposes an energy minimization framework to encompass these three problems. This framework should provide researchers with a thorough understanding of a large number of existing time series techniques, and it may serve as a tool for dealing with other problems in time series as they arise.Enabling computers to understand human behavior has the potential to revolutionize many areas that benefit society such as clinical diagnosis, human computer interaction, and social robotics. Advances in time series to model human actions and events from sensory data have been critical to the success of systems that can recognize and characterize human behavior. However, most existing algorithms have been supervised in nature. Supervised learning typically requires large amounts of human annotation, that is typically labor intensive and is difficult to standardize across coders. In this proposal the PI explores the use of unsupervised learning techniques for aligning and discovering patterns in time series of human motion that have been captured with accelerometers, video or motion capture technologies. The PI will show how the proposed algorithms outperform state-of-the-art techniques in several human sensing tasks such as temporal alignment of human motion, temporal clustering of human activities from video, learning motion primitives, and joint segmentation and classification of human behavior. In the educational aspect, the PI will continue to provide support to the Carnegie Science Center to demonstrate human sensing technologies, as well as incorporate a large number of undergraduates in his research. The research source code will be made available to the scientific community.