Dataset for Human Activity Recognition

Recognizing actions through visual cues poses a particularly formidable challenge within the domains of computer vision and pattern recognition. A specific practical application of this challenge involves the swift identification of instances of physical altercations, such as fights, captured by surveillance cameras in public spaces and correctional facilities. Nevertheless, the field of action recognition has primarily focused its efforts on relatively straightforward actions, such as clapping, walking, wrestling, and jogging. In contrast, the identification of specific events with immediate practical applications, such as cellphone snatching, common fighting, and running behavior in general, has received comparatively less attention. The ability to detect such events could prove immensely valuable in various video surveillance contexts, including prisons, psychiatric facilities, and violence detection based on human activity. As a result, there is a growing interest in the development of algorithms and gold-standard benchmarks geared toward detecting instances of violence and abnormal behaviors based on human activity recognition problems. To address this limitation, this study presents a substantial benchmark dataset comprising of totaling 6,048 video frames, featuring 2016 cases of cellphone snatching, 2016 cases of fighting, and 2016 cases of running, each showcasing five different forms of human activities. Our proposed dataset will be made publicly available to foster and promote research in human action recognition behaviors, including the development of robbery detection systems, human movement detection systems, safety systems, theft detection systems, and anomaly detection in automatic surveillance cameras.

借助视觉线索进行动作识别，在计算机视觉与模式识别领域堪称极具挑战性的难题。该难题的一项具体实际应用场景，便是快速识别公共场所与惩戒机构监控摄像头捕捉到的肢体冲突事件，例如斗殴行为。然而，当前动作识别领域的研究主要聚焦于较为简单的动作类别，如鼓掌、行走、摔跤与慢跑等。与之形成对比的是，那些具备直接实用价值的特定事件识别——如抢夺手机、一般性斗殴与常规奔跑行为——却相对较少受到研究关注。检测这类事件的能力，在监狱、精神卫生机构等各类视频监控场景中，以及基于人类活动的暴力检测场景里，都将展现出极高的应用价值。因此，针对基于人类活动识别的暴力与异常行为检测任务，开发相关算法与金标准基准数据集的研究兴趣与日俱增。为解决这一局限，本研究构建了一个规模可观的基准数据集，共计包含6048个视频帧，涵盖2016例抢夺手机事件、2016例斗殴事件以及2016例奔跑行为，每类事件均包含五种不同的人类活动表现形式。本研究提出的数据集将公开发布，以推动与促进人类动作识别相关研究，包括抢劫检测系统、人体运动检测系统、安全防护系统、盗窃检测系统以及自动监控摄像头异常检测等方向的研发工作。