基于Myo臂环的手势动作数据集

基于Myo臂环的手势动作数据集是为构建手势识别模型的训练基础数据，识别模型能够将识别结果与现场大规模荧光棒实现互动控制，实现基于可穿戴设备的动作识别与交互视觉呈现。Myo臂环可以采集干电极采集到的皮肤表面由于肌肉收缩产生的生物电信号，还可以通过臂环惯性测量单元测量到的手臂运动时的运动信号，两者都可以反映手部的运动信息。我们建立了包含20类手势动作的数据集，其中包括9种动态手势动作和11种静态手势动作，动态手势动作包括6个方向的平移动作和3个坐标平面的旋转动作，静态手势动作包括1个握拳动作和0~9这10个数字动作。由于不同类型的手势动作的实际执行时间存在差异，为保证手势动作数据的有效性，仅保留了各类动作的有效长度。同时剔除了数据集中的错误数据，建立了包含19729个手势动作数据的数据集。人工对数据活动段的起点和终点进行标注，注明数据段的手势动作类别。使用滑动窗口将手势动作信号分割为固定长度的、有一定数据重叠的信号段作为交互系统的实时控制信号。滑动时间窗口长度为1s、步长为100ms。

The Myo armband-based gesture dataset serves as training data for building gesture recognition models, which can interactively control large-scale on-site light sticks using recognition results, achieving wearable device-based motion recognition and interactive visual presentation. The Myo armband can collect bioelectrical signals on the skin surface generated by muscle contractions via dry electrodes, as well as motion signals during arm movement measured by the armband's inertial measurement unit (IMU); both reflect hand motion information. We constructed a dataset containing 20 categories of gestures, including 9 dynamic gestures and 11 static gestures. The dynamic gestures cover 6 directional translation movements and rotation movements in 3 coordinate planes, while the static gestures include 1 fist gesture and 10 numeric gestures ranging from 0 to 9. Given that the actual execution durations of different gesture types vary, to ensure the validity of gesture data, only the valid segments of each gesture were retained, and erroneous data in the dataset were removed, resulting in a dataset containing 19729 gesture samples. The start and end points of active data segments were manually annotated, with the gesture category of each segment labeled. Sliding windows were used to split the gesture motion signals into fixed-length, overlapping signal segments as real-time control signals for the interactive system. The sliding time window has a length of 1 second and a stride of 100 ms.

搜集汇总

背景与挑战

背景概述

该数据集是基于Myo臂环采集的手势动作数据，包含20类手势动作（9种动态和11种静态），用于手势识别模型的训练。数据集经过清洗和标注，包含19729个有效手势动作数据，总大小为1.77GB。