Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Anomaly detection in non-stationary videos using time-recursive differencing network based prediction

Gargi V. Pillai, Debashis Sen

TL;DR

The differencing network is employed to effectively handle nonstationarity in video data during the anomaly detection and comparison with several existing methods including the state-of-the-art reveal the superiority of the proposed approach.

Abstract

Most videos, including those captured through aerial remote sensing, are usually non-stationary in nature having time-varying feature statistics. Although, sophisticated reconstruction and prediction models exist for video anomaly detection, effective handling of non-stationarity has seldom been considered explicitly. In this paper, we propose to perform prediction using a time-recursive differencing network followed by autoregressive moving average estimation for video anomaly detection. The differencing network is employed to effectively handle non-stationarity in video data during the anomaly detection. Focusing on the prediction process, the effectiveness of the proposed approach is demonstrated considering a simple optical flow based video feature, and by generating qualitative and quantitative results on three aerial video datasets and two standard anomaly detection video datasets. EER, AUC and ROC curve based comparison with several existing methods including the state-of-the-art reveal the superiority of the proposed approach.

Anomaly detection in non-stationary videos using time-recursive differencing network based prediction

TL;DR

The differencing network is employed to effectively handle nonstationarity in video data during the anomaly detection and comparison with several existing methods including the state-of-the-art reveal the superiority of the proposed approach.

Abstract

Most videos, including those captured through aerial remote sensing, are usually non-stationary in nature having time-varying feature statistics. Although, sophisticated reconstruction and prediction models exist for video anomaly detection, effective handling of non-stationarity has seldom been considered explicitly. In this paper, we propose to perform prediction using a time-recursive differencing network followed by autoregressive moving average estimation for video anomaly detection. The differencing network is employed to effectively handle non-stationarity in video data during the anomaly detection. Focusing on the prediction process, the effectiveness of the proposed approach is demonstrated considering a simple optical flow based video feature, and by generating qualitative and quantitative results on three aerial video datasets and two standard anomaly detection video datasets. EER, AUC and ROC curve based comparison with several existing methods including the state-of-the-art reveal the superiority of the proposed approach.

Paper Structure

This paper contains 9 sections, 6 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Proposed Prediction Approach for Video Anomaly Detection
  3. Motivation
  4. Time-recursive Differencing based Anomaly Detection
  5. Model and Hyper-Parameters
  6. Experimental results
  7. Results and Discussion
  8. Conclusion
  9. Aknowledgement

Figures (5)

  • Figure 1: Schematic of the proposed approach, where the time-recursive differencing module handles non-stationary inputs to produce stationary inputs for anomaly detection. $f(A_{m,n})$ refers to feature extracted from the $m^{th}$ block of $n^{th}$ frame.
  • Figure 2: Anomalies detected in frames from aerial videos, (a) walking people suddenly start running in a video from the UCF-AA dataset and (b) a cyclist appears in the scene in a video from the SD dataset
  • Figure 3: Anomalies detected in frames from the UCSD Ped2 dataset, (a) vehicle and (b) skateboarding in pedestrian path
  • Figure 4: Anomalies detected in consecutive frames of videos from (a) SD, (b) OA, (c)-(d) UCF-AA, (e) UCSD Ped2 and (f)-(h) UMN datasets
  • Figure 5: Frame-level ROC curves for (a) UCSD Ped2, and (b) UMN datasets.