基于NvGesture的手势识别研究数据集

文中提出的多模态动态手势数据集（multi-modal dynamic hand gesture dataset），主要面向动态手势识别算法研究与优化人机交互需求建设。基于汽车模拟器环境，利用 SoftKinetic DS325 传感器采集前视图的颜色和深度视频，通过顶部安装的 DUO 3D 传感器记录一对立体红外流。该数据集的产生方法是：邀请 20 名受试者在汽车模拟器中，左手控制方向盘、右手做手势，界面会随机提示 25 种不同的手势让受试者操作，每种手势重复 3 次。这些手势包括手指和手部的多种动作，如移动、点击、招手等。数据集主要内容涵盖了不同光照条件下的连续数据流，包含 1532 个动态手势，采集的模态丰富，除上述传感器数据外，还计算了密集光流和红外视差图。其意义在于有效弥补现有手势数据集的不足，为评估和改进手势识别算法提供了高质量的测试基准，推动人机交互领域的技术发展。

The multi-modal dynamic hand gesture dataset proposed in this paper is constructed primarily for the research of dynamic hand gesture recognition algorithms and the optimization of human-computer interaction (HCI) requirements. Built upon a car simulator environment, it collects color and depth videos from the front view using a SoftKinetic DS325 sensor, and records a pair of stereo infrared streams via a top-mounted DUO 3D sensor. The dataset is generated as follows: 20 subjects are invited to operate in the car simulator, with their left hand controlling the steering wheel and the right hand performing gestures. The interface randomly prompts 25 distinct gestures for the subjects to execute, and each gesture is repeated 3 times. These gestures cover various finger and hand movements such as moving, tapping, waving, and so on. The main content of the dataset includes continuous data streams under different lighting conditions, with a total of 1532 dynamic gestures. It features rich acquisition modalities: in addition to the aforementioned sensor data, dense optical flow and infrared disparity maps are also calculated. Its significance lies in effectively compensating for the shortcomings of existing hand gesture datasets, providing a high-quality test benchmark for evaluating and improving gesture recognition algorithms, and promoting technological development in the human-computer interaction field.

搜集汇总

背景与挑战

背景概述

该数据集是一个多模态动态手势数据集，专为手势识别算法研究和人机交互优化而设计。它采集自汽车模拟器环境，包含颜色、深度、立体红外流等多种传感器数据，总计1532个手势，涵盖25种不同手势类型，并在不同光照条件下录制，以提供高质量的测试基准，推动技术发展。