UC2017 Static and Dynamic Hand Gestures

We introduce the UC2017 static and dynamic gesture dataset. Most researchers use vision-based systems such as the Microsoft Kinect to acquire and classify hand gesture data. Despite that, we believe that we can achieve more reliable results and allow the use of more complex gestures with wearable systems. There are not many datasets with wearable systems due to the plethora of data gloves in the market and their relative high cost. For these reasons, we opted by creating a new dataset to present and evaluate our gesture recognition framework. The objectives of the dataset are: (1) provide a superset of hand gestures for HRI, (2) have user variability, (3) to be representative of the actual gestures performed in a real-world interaction. We divide the dataset in two types of gestures: SG and DG. SG are described by a single timestep of data, therefore representing a single hand pose and orientation. DGs are variable-length timeseries of poses and orientations with particular meanings. Some of the gestures of the dataset are correlated with a certain meaning in the context of HRI, while others are arbitrary, to enrich the dataset and add complexity to the classification problem. The library is composed of 24 SG classes and 10 DG. The dataset includes SG data from eight subjects with a total of 100 repetitions for each of the 24 classes (2400 samples in total). The DG samples were obtained from six subjects and has cumulatively 131 repetitions of each class (1310 samples in total). All of the subjects are right-handed and performed the gestures with their left hand. We used a data glove (CyberGlove II) and a magnetic tracker (Polhemus Liberty) to capture the hand shape, position and orientation over time. The glove provides digital signals that are proportional to the bending angle of each one of the 22 sensors which are elastically attached to a subset of the hand's joints. In this way we have an approximation of the hand's shape. The tracker's sensor is rigidly attached to the glove on the wrist and measures its position and orientation in respect to a ground-fixed frame. The orientation is the rotation between the fixed frame and the frame of the sensor, given a quaternion (WXYZ). We fuse the sensor data together online since the sensors have slightly different acquisition rates -- 100Hz for the glove and 120Hz for the tracker. The tracker data are under-sampled by gathering only the closest tracker frame in time. The files are in the h5df format. The dimensions are (sample, time, variables).

我们提出了UC2017静态与动态手势数据集。当前多数研究者采用微软Kinect（Microsoft Kinect）这类基于视觉的系统来获取并分类手势数据。尽管如此，我们认为借助可穿戴系统能够获得更可靠的识别结果，同时支持更为复杂的手势应用。由于市面上数据手套品类繁杂且成本相对高昂，可穿戴系统相关的手势数据集较为稀缺。基于上述原因，我们构建了全新的数据集，用于展示并评估我们的手势识别框架。

本数据集的构建目标如下：(1) 为人类-机器人交互（Human-Robot Interaction, HRI）场景提供覆盖全面的手势集合；(2) 涵盖不同使用者的个体差异；(3) 能够真实反映实际交互场景中的手势行为。

我们将数据集内的手势分为两类：静态手势（Static Gesture, SG）与动态手势（Dynamic Gesture, DG）。静态手势仅包含单个时间步的数据，对应单一的手部姿态与朝向；动态手势则为带有特定语义、长度可变的姿态与朝向时序序列。本数据集中部分手势对应人类-机器人交互场景下的特定语义，其余则为随机设计的手势，以此丰富数据集的多样性并提升分类任务的复杂度。

该数据集包含24个静态手势类别与10个动态手势类别。其中静态手势数据来自8名受试者，24个类别各重复采集100次，总计2400个样本。动态手势数据则来自6名受试者，每个类别累计采集131次，总计1310个样本。所有受试者均为右利手，且均使用左手完成手势采集。

我们采用数据手套（data glove）CyberGlove II与磁性跟踪器（magnetic tracker）Polhemus Liberty来实时采集手部形状、位置与朝向数据。该数据手套搭载22个弹性绑定于手部部分关节的传感器，可输出与各传感器弯曲角度成正比的数字信号，以此实现手部形状的近似还原。跟踪器的传感器被刚性固定于手套的腕部，可测量传感器相对于地面固定坐标系的位置与朝向。朝向信息以四元数（WXYZ）的形式表示，对应固定坐标系与传感器坐标系之间的旋转变换。

由于两类传感器的采集速率略有差异（数据手套为100Hz，跟踪器为120Hz），我们在线对传感器数据进行融合处理。针对跟踪器数据，我们通过仅选取时间上最邻近的跟踪帧的方式进行下采样。数据集文件采用h5df格式存储，数据维度为（样本数、时间步长、变量数）。

搜集汇总

背景与挑战

背景概述

UC2017 Static and Dynamic Hand Gestures数据集是一个用于人机交互研究的手势数据集，包含24个静态手势类别和10个动态手势类别，通过CyberGlove II数据手套和Polhemus Liberty磁性追踪器采集手部形状、位置和方向数据。数据集共有3710个样本，涉及多名受试者，旨在提供可靠的手势识别资源。