Abstract
Detecting the anomaly of human behavior is paramount to timely recognizing endangering situations, such as street fights or elderly falls. However, anomaly detection is complex, since anomalous events are rare and because it is an open set recognition task, i.e., what is anomalous at inference has not been observed at training. We propose COSKAD, a novel model which encodes skeletal human motion by an efficient graph convolutional network and learns to COntract SKeletal kinematic embeddings onto a latent hypersphere of minimum volume for Anomaly Detection. We propose and analyze three latent space designs for COSKAD: the commonly-adopted Euclidean, and the new spherical-radial and hyperbolic volumes. All three variants outperform the state-of-the-art, including video-based techniques, on the ShangaiTechCampus, the Avenue, and on the most recent UBnormal dataset, for which we contribute novel skeleton annotations and the selection of human-related videos. The source code and dataset will be released upon acceptance.
Abstract (translated)
检测人类行为异常是及时识别危险情况的关键,例如街头打架或老年人摔倒。然而,异常检测是复杂的,因为异常事件罕见,而且它是开放集识别任务,即推断出异常行为在训练时没有被观察到。我们提出了COSKAD,一个新模型,通过高效的图形卷积网络编码骨骼人类运动,并学习将骨骼运动嵌入到最小体积的隐态Hyper球上进行异常检测。我们提出了并分析三个 COSKAD 的隐态空间设计:常见的欧几里得空间,和新开发的球形Radial 和Hyper空间。所有三个变体在ShanghaiTech Campus、Ave和最近发布的Ubnormal数据集上表现更好,我们为这些数据集提供了新的骨骼注释和选择与人类相关的视频。源代码和数据集将在接受后发布。
URL
https://arxiv.org/abs/2301.09489