A Dynamic Hand Gesture Recognition Dataset for Human-Computer Interfaces

Computer vision systems are commonly used to design touch-less human-computer interfaces (HCI) based on dynamic hand gesture recognition (HGR), which have a wide range of applications in several domains, such as, gaming, multimedia, automotive, home automation. However, automatic HGR is still a challenging task, mostly because of the diversity in how people perform the gestures. In addition, the number of publicly available hand gesture datasets is scarce and often the gestures are not acquired with sufficient image quality, and the gestures are not correctly performed.For these reasons, we propose a dataset of 27 dynamic hand gesture types acquired at full HD resolution from 21 different subjects, which were carefully instructed before performing the gestures and were observed during the gesture performing; in case the performed hand gesture was not correct, the subjects had to repeat the movement. Each subject performed 3 times the 27 hand gestures for a total of 1701 videos collected and corresponding 204120 video frames.

计算机视觉系统通常被用于基于动态手势识别（Hand Gesture Recognition, HGR）构建无接触式人机交互界面（Human-Computer Interaction, HCI），该类界面在游戏、多媒体、车载、家居自动化等多个领域拥有广泛应用。然而，自动手势识别仍属于一项极具挑战性的任务，其核心难点在于不同用户的手势动作存在显著差异。此外，当前公开可用的手部手势数据集数量稀少，且多数数据集存在图像质量不足、手势动作未正确执行等问题。基于上述原因，本研究提出了一个包含27种动态手势类别的数据集：该数据集以全高清（Full HD）分辨率采集自21名不同受试者，所有受试者在执行手势前均经过了细致的动作指导，并在手势执行过程中受到监督；若受试者的手势动作未达标，则需重新执行该动作。每名受试者需完成27种手势各3次，最终共采集得到1701段视频，对应总计204120帧图像。