Scene perception guided crowd anomaly detection

Xuguang Zhang, Dingxin Ma, Hui Yu*, Ya Huang, Peter Howell, Brett Stevens

*Corresponding author for this work

    Research output: Contribution to journalArticlepeer-review

    404 Downloads (Pure)

    Abstract

    Crowd anomaly detection has been a research hotspot in the field of video surveillance in recent years. In most existing methods, the accuracy of anomaly detection dominantly relies on the acquisition of regions of interest (ROI) and feature extraction. However, the randomness of ROI segmentation and crowd group selection usually cannot guarantee a robust performance and thus may lead to false detection sometimes. To address these issues, this paper proposes a scene perception-based approach combining the fluid forces expression and psychological theory. The proposed method firstly introduces a flow field visualization technology called line integral convolution to segment the moving pedestrians in the scene. Then, a scene perception-guided clustering strategy is proposed to cluster the consistency crowd group. Scene perception strategy is in line with the psychological criteria of human cognition. In clustering, it makes more reasonable use of various attributes of the pedestrians. To ensure a robust detection of the pedestrian group, we propose a fluid feature concept which considers both mass force and surface force. For each consistency group, two types of features including the image appearance feature and fluid feature are combined to describe pedestrian motion. The experimental results show that the proposed method achieves higher accuracy in comparison with some existing methods in terms of both frame-level and pixel-level measurements.
    Original languageEnglish
    Article number0
    Pages (from-to)291-302
    Number of pages12
    JournalNeurocomputing
    Volume414
    Early online date30 Jul 2020
    DOIs
    Publication statusPublished - 13 Nov 2020

    Fingerprint

    Dive into the research topics of 'Scene perception guided crowd anomaly detection'. Together they form a unique fingerprint.

    Cite this