Berkeley MHAD (Berkeley Multimodal Human Action Database)

Berkeley Multimodal Human Action Database (MHAD) 包含 7 名男性和 5 名女性受试者在 23-30 岁之间执行的 11 项动作，除了一名老年受试者。所有受试者每个动作重复 5 次，产生大约 660 个动作序列，对应于大约 82 分钟的总记录时间。此外，我们为每个主题记录了一个 T-pose，可用于骨骼提取；和背景数据（在某些活动中使用和不使用椅子）。图 1 显示了前置摄像头所采取的所有动作的快照以及从 Kinect 数据中提取的相应点云。指定的一组动作包括以下内容：(1) 上下肢运动的动作，例如原地跳跃、起重跳、投掷等，(2) 上肢高动态的动作，例如，挥手、拍手等和（3）下肢高动态动作，例如坐下、站起来。在每次录制之前，都会向受试者指示要执行什么动作；但是没有给出关于如何执行动作的具体细节（即性能风格或速度）。因此，受试者在执行某些动作（例如，拳击、投掷）时采用了不同的风格。图 2 显示了来自每个摄像头集群的参考摄像头和两个 Kinect 摄像头的投掷动作快照。该图展示了与单视点相比，可以从多视图和深度观察中获得的信息量。

Berkeley Multimodal Human Action Database (MHAD) contains 11 action categories performed by a total of 13 subjects: 7 males and 5 females aged 23–30 years, plus one elderly subject. All subjects repeated each action 5 times, generating approximately 660 action sequences with a total recorded duration of around 82 minutes. Additionally, a T-pose was recorded for each subject, which can be utilized for skeleton extraction, alongside background data (collected with and without a chair during certain activities). Figure 1 displays snapshots of all actions captured by the front-facing camera, along with the corresponding point clouds extracted from the Kinect data. The specified set of actions is divided into three groups: (1) actions involving movements of both upper and lower extremities, such as jumping in place, weighted jump, throwing, and others; (2) actions featuring high dynamic movements of the upper extremities, such as waving hands, clapping hands, and others; and (3) actions featuring high dynamic movements of the lower extremities, such as sitting down and standing up. Prior to each recording session, subjects were informed of the target action to perform, but no specific guidelines regarding execution details (i.e., performance style or speed) were provided. As a result, subjects adopted varied execution styles for certain actions such as boxing and throwing. Figure 2 presents snapshots of the throwing action captured by the reference camera in each camera cluster and the two Kinect cameras. This figure demonstrates the increased amount of information that can be acquired from multi-view and depth observations compared to a single viewpoint.

搜集汇总

背景与挑战

背景概述

Berkeley MHAD数据集包含12名受试者（7男5女）执行的11项动作，每个动作重复5次，共约660个动作序列，总记录时间约82分钟。数据集涵盖了上下肢运动、上肢高动态和下肢高动态动作，适用于多模态动作分析研究。