Hockey Fight Detection Dataset

虽然动作识别社区主要关注检测拍手、走路或慢跑等简单动作，但对打架或一般攻击性行为的检测研究相对较少。这种功能在某些视频监控场景中可能非常有用，例如监狱、精神病院或老年中心，甚至在照相手机中。在分析了以前的方法之后，我们测试了用于在特定的战斗检测问题中进行动作识别的著名词袋框架，以及目前可用的两个最佳动作描述符：STIP 和 MoSIFT。为了评估和促进视频中暴力检测的研究，我们引入了一个新的视频数据库，其中包含 1000 个序列，分为两组：打架和非打架。对该数据库和另一个与动作电影打斗的数据库进行的实验表明，可以以接近 90% 的准确率检测到打斗。

While the action recognition community has largely focused on detecting simple actions such as clapping, walking, or jogging, research on detecting fighting or general aggressive behaviors has received relatively little attention. This functionality could be highly useful in certain video surveillance scenarios, such as prisons, psychiatric hospitals, or senior care centers, and even on camera phones. After analyzing prior approaches, we tested the well-known bag-of-words framework for action recognition on the specific fighting detection task, along with the two best currently available action descriptors: STIP and MoSIFT. To evaluate and advance research on violence detection in videos, we introduce a novel video database containing 1000 video sequences, divided into two categories: fighting and non-fighting. Experiments conducted on both this database and another database focused on action movie fight scenes demonstrate that fighting can be detected with near 90% accuracy.