Multi-scale and real-time non-parametric approach for anomaly detection and localization

https://doi.org/10.1016/j.cviu.2011.09.009Get rights and content

Abstract

In this paper we propose an approach for anomaly detection and localization, in video surveillance applications, based on spatio-temporal features that capture scene dynamic statistics together with appearance. Real-time anomaly detection is performed with an unsupervised approach using a non-parametric modeling, evaluating directly multi-scale local descriptor statistics. A method to update scene statistics is also proposed, to deal with the scene changes that typically occur in a real-world setting. The proposed approach has been tested on publicly available datasets, to evaluate anomaly detection and localization, and outperforms other state-of-the-art real-time approaches.

Highlights

► In this paper we present a non-parametric approach to anomaly detection in surveillance videos. ► The real-time system uses spatio-temporal features, integrated in a multi-scale approach. ► The system can localize anomalies temporally (at frame level) and spatially (within frame). ► The systems has been compared to state-of-the-art approaches on a real-world UCSD dataset. ► According to experiments our method consistently outperforms other real-time approaches.

Section snippets

Introduction and previous work

The real-world surveillance systems currently deployed are primarily based on the performance of human operators that are expected to watch, often simultaneously, a large number of screens (up to 50 [2]) that show streams captured by different cameras. One of the main tasks of security personnel is to perform proactive surveillance to detect suspicious or unusual behavior and individuals [3] and to react appropriately. As the number of CCTV streams increases, the task of the operator becomes

Scene representation

Modeling crowd patterns is one of the most complex contexts for detection of anomalies in video surveillance scenarios. Describing such statistics is extremely complex since, as stated in Section 1, the use of trajectories does not allow to capture all the possible anomalies that may occur, e.g. due to variations of scene appearance and the presence of unknown objects moving in the scene; this is due to the fact that object detection and tracking are often unfeasible both for computational

Real-time anomaly detection

Our system is able to learn from a normal data distribution fed as a training set but can also start without any knowledge of the scene, learning and updating the “normal behavior” profile dynamically, almost without any human intervention. The model can always be updated with a very simple procedure. Despite the simple formulation of this approach our system is able to model complex and crowded scenes, including both dynamic and static patterns.

Our technique is inspired by the one proposed in

Experimental results

We tested our approach on the UCSD1 anomaly dataset presented in [34], which provides frame-by-frame local anomaly annotation. The dataset consists of two subsets, corresponding to different scenes using fixed cameras that overlook pedestrian walkways: one (called Peds1) contains videos of people moving towards and away from the camera, with some perspective distortion; the other (called Peds2) shows pedestrian movement parallel to the

Conclusions

In this paper we have presented a multi-scale non-parametric anomaly detection approach that can be executed in real-time in a completely unsupervised manner. The approach is capable of localizing anomalies in space and time. We have also provided a straightforward procedure to dynamically update the learned model, to deal with scene changes that happen in real-world surveillance scenarios. Dense and overlapping spatio-temporal features, that model appearance and motion information, have been

References (42)

  • V. Chandola et al.

    Anomaly detection: a survey

    ACM Comput. Surv.

    (2009)
  • C. Brax, L. Niklasson, M. Smedberg, Finding behavioural anomalies in public areas using video surveillance data, in:...
  • I. Ivanov, F. Dufaux, T.M. Ha, T. Ebrahimi, Towards generic detection of unusual events in video surveillance, in:...
  • S. Calderara, C. Alaimo, A. Prati, R. Cucchiara, A real-time system for abnormal path detection, in: Proc. of 3rd IEE...
  • J. Li, S. Gong, T. Xiang, Discovering multi-camera behaviour correlations for on-the-fly global activity prediction and...
  • C. Liu, G. Wang, W. Ning, X. Lin, L. Li, Z. Liu, Anomaly detection in surveillance video using motion direction...
  • C. Piciarelli et al.

    Surveillance-oriented event detection in video streams

    IEEE Intelligent Systems

    (2011)
  • D. Zhang, D. Gatica-Perez, S. Bengio, I. McCowan, Semi-supervised adapted HMMs for unusual event detection, in: Proc....
  • R. Sillito, R. Fisher, Semi-supervised learning for anomalous trajectory detection, in: Proc. of British Machine Vision...
  • O. Boiman et al.

    Detecting irregularities in images and in video

    Int. J. Comput. Vision (IJCV)

    (2007)
  • A. Adam et al.

    Robust real-time unusual event detection using multiple fixed-location monitors

    IEEE Trans. Pattern Anal. Mach. Intell. (PAMI)

    (2008)
  • Cited by (0)

    An earlier version of this paper appeared in a conference proceeding [1].

    View full text