Contrasting Estimation of Pattern Prototypes for Anomaly Detection in Urban Crowd Flow
Yupeng Wang, Xiling Luo, Zequan Zhou
Abstract
Crowd flow anomaly detection (CFAD) plays a vital role in ensuring public safety due to its capability to distinguish abnormal crowd movement behaviors from the norm. However, the influence of coexisting spatiotemporal factors presents a substantial challenge in capturing the dynamic normal pattern. Moreover, the inherent similarities within the original crowd flow data necessitate the creation of a discriminative feature space. To address this, we propose <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ProtoDetect</i> , a novel method that learns distinguishable representations (named as prototypes) of the influencing factors. It subsequently identifies anomalous samples by comparing them with their normal counterparts, estimated based on the prototypes. Experimental evaluations on three real-world datasets demonstrate ProtoDetect’s consistent superior performance in CFAD. The source code is available at https://github.com/yupwang/ProtoDetect.