SBU Kinect Interaction 人体肢体动作视频数据

使用身体姿势特征和多实例学习的双人交互检测 人类活动识别有可能影响从监视到人机界面到基于内容的视频检索的广泛应用。最近，便宜的深度传感器（例如，Microsoft Kinect）的快速发展为活动识别应用的实时全身人体跟踪提供了足够的准确性。在本文中，我们创建了一个复杂的人类活动数据集，描述了两个人的交互，包括同步视频，深度和动作捕捉数据。此外，在通过支持向量机（SVM）实时检测交互活动的上下文中，我们使用数据集来评估通常用于索引和检索运动捕获数据的各种功能。实验上，我们发现基于所有关节对之间距离的几何关系特征优于其他特征选择。对于整个序列分类，我们还探索了与多实例学习（MIL）相关的技术，其中序列由一袋身体姿势特征表示。我们发现当序列在感兴趣的相互作用周围时间延伸时，基于MIL的分类器优于SVM。 在第二届计算机视觉与模式识别会议上的3D数据人类活动理解国际研讨会上发表，CVPR 2012 Kiwon Yun，Jean Honorio，Debaleena Chattopadhyay，Tamara L. Berg，Dimitris Samaras Stony Brook University

Two-Person Interaction Detection Using Body Pose Features and Multiple Instance Learning

Human activity recognition has the potential to impact a wide range of applications spanning surveillance, human-computer interfaces and content-based video retrieval. Recently, the rapid development of low-cost depth sensors (e.g., Microsoft Kinect) has provided sufficient accuracy for real-time full-body human tracking in activity recognition applications. In this paper, we create a complex human activity dataset depicting interactions between two people, which includes synchronized video, depth and motion capture data. Furthermore, in the context of real-time interaction activity detection using Support Vector Machines (SVMs), we utilize the dataset to evaluate various features commonly used for indexing and retrieving motion capture data. Experimentally, we find that geometric relationship features based on pairwise distances between all joint pairs outperform other feature selection strategies. For full sequence classification, we also explore techniques related to Multiple Instance Learning (MIL), where a sequence is represented as a bag of body pose features. We observe that MIL-based classifiers outperform SVMs when sequences are temporally extended around the interaction of interest.

Published in the 2nd International Workshop on Human Activity Understanding from 3D Data, co-located with the Conference on Computer Vision and Pattern Recognition (CVPR 2012)
Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L. Berg, Dimitris Samaras
Stony Brook University

搜集汇总

背景与挑战

背景概述

SBU Kinect Interaction数据集是一个包含同步视频、深度和动作捕捉数据的人体肢体动作视频数据集，专注于双人交互检测和人类活动识别研究。该数据集适用于多种应用场景，如监视和人机界面，并采用了身体姿势特征和多实例学习技术来优化交互活动的检测。