PaperlayerView exact revision on arXiv

Source-linked AI summary

Online Anomaly Detection in Surveillance Videos with Asymptotic Bounds on False Alarm Rate

Keval Doshi, Yasin Yilmaz

arXiv:2010.07110v1cs.CVcs.LGeess.IVstat.ML

TL;DR

Online video anomaly detection is important for timely decisions, but remains relatively little studied. MONAD combines neural feature extraction with statistical sequential detection and a threshold-selection procedure based on an asymptotic false-alarm bound. The authors report outperforming state-of-the-art methods and false-alarm rates consistent with the derived bounds.

  • Problem

    Online decision making is important for timely responses to abnormal events, but relatively little research has focused on online and real-time video anomaly detection.

    Source: §1 · p. 2

  • Method

    MONAD combines deep learning-based feature extraction with a statistical sequential anomaly detector and derives an asymptotic false-alarm bound for threshold selection.

    Source: §1 · p. 3

  • Results

    MONAD outperforms state-of-the-art methods and yields false-alarm rates consistent with its derived asymptotic bounds in real-data experiments.

    Source: §5 · p. 18

  • Takeaways & Limitations

    Given the training data, the procedure makes it possible to compute a threshold intended to achieve a desired false-alarm rate asymptotically.

    Source: §3.7 · p. 13

Abstract

from arXiv · show

Anomaly detection in surveillance videos is attracting an increasing amount of attention. Despite the competitive performance of recent methods, they lack theoretical performance analysis, particularly due to the complex deep neural network architectures used in decision making. Additionally, online decision making is an important but mostly neglected factor in this domain. Much of the existing methods that claim to be online, depend on batch or offline processing in practice. Motivated by these research gaps, we propose an online anomaly detection method in surveillance videos with asymptotic bounds on the false alarm rate, which in turn provides a clear procedure for selecting a proper decision threshold that satisfies the desired false alarm rate. Our proposed algorithm consists of a multi-objective deep learning module along with a statistical anomaly detection module, and its effectiveness is demonstrated on several publicly available data sets where we outperform the state-of-the-art algorithms. All codes are available at https://github.com/kevaldoshi17/Prediction-based-Video-Anomaly-Detection-.

1. Introduction

The section frames online anomaly detection and performance guarantees as under-addressed needs in surveillance video analysis. It introduces a hybrid neural and statistical framework, an asymptotic false-alarm bound, and a threshold-selection procedure.

1. Introduction

  • Online decision making is important for timely responses to abnormal events, yet relatively little research addresses online detection and some nominally online methods rely on batch processing.One cited example requires an entire video for normalization.Source: §1 · pp. 2–3
  • The authors identify a performance-analysis gap in recent methods that rely on complex neural networks, while noting that statistical and nearest-neighbor approaches are amenable to analysis.They also cite concerns about interpretability, analyzability, and reliability of neural-network decisions.Source: §1 · p. 3
  • The proposed framework combines deep learning-based feature extraction with a statistical sequential anomaly detector for online video anomaly detection.The authors also describe the approach as a hybrid of neural networks and statistical k-nearest-neighbor decision making.Source: §1 · p. 3
  • The method derives an asymptotic false-alarm bound and provides a procedure for choosing a threshold that meets a desired false-alarm rate.The authors motivate this procedure by noting that existing work treats the threshold as a design parameter without an analytical selection procedure.Source: §1 · p. 3
  • The framework is extensively evaluated on publicly available video anomaly detection datasets.Source: §1 · p. 3

2. Related Work

Prior work spans semi-supervised and supervised learning, hand-crafted or trajectory-based features, and neural-network feature learning. The section notes limits tied to anomaly labels, feature scope, and unfamiliar anomaly types.

2. Related Work

  • Semi-supervised methods learn normality from nominal videos, while supervised methods use nominal and anomalous videos but can struggle with representative frame-level labels and novel anomalies.Weakly supervised multiple-instance learning uses video-level annotations.Source: §2 · p. 4
  • Trajectory-focused features can restrict detection to anomalies involving trajectories or moving objects, such as abnormal vehicle paths or human-skeleton patterns.Source: §2 · p. 4
  • Other hand-crafted approaches use motion and appearance features, optical-flow or oriented-gradient histograms, and sparse coding or reconstruction.The cited sparse-coding methods model normal events or joint object trajectories.Source: §2 · pp. 4–5
  • Neural-network approaches learn appearance and motion features, including through CNN/CLSTM models, GAN-based methods, and transfer learning [29].Source: §2 · p. 5

3. Proposed Method

MONAD combines deep feature extraction with nonparametric sequential detection to make online video-anomaly decisions and derive a threshold for a desired false alarm rate. Its features cover appearance, motion, and location, while its detector models nominal data using k-nearest-neighbor distances.

3.1. Motivation

  • MONAD unifies a GAN-based future-frame predictor and lightweight YOLOv3 object detector with a nonparametric statistical algorithm for online anomaly detection.The framework is presented as providing theoretical performance analysis for a deep-learning-based video anomaly detector.Source: §3.1 · p. 6

3.2. Feature Selection

  • Appearance, motion, and location are considered because anomalies may be expressed through any of these aspects, and the feature categories can be weighted by relative importance.The feature vector uses weights w1, w2, and w3 for motion, location, and appearance, respectively.Source: §3.2 · p. 7

3.3. Frame Prediction

  • Future-frame prediction error supplies motion features, and the predictor uses a Least Square GAN with intensity, gradient-difference, optical-flow, and adversarial training constraints.The generator uses a U-Net, while the discriminator is a patch discriminator; the losses are weighted in its objective.Source: §3.3 · pp. 7–8

3.4. Object Detection

  • YOLOv3 supplies object bounding-box location and class-probability appearance features, with box centers and areas used to represent location.The text favors YOLOv3 for implementation because speed is critical for online detection; centers and areas also reduce complexity and can avoid false positives from loose boxes.Source: §3.4 · p. 9

3.5. Feature Vector

  • For each detected object, the feature vector combines prediction MSE, bounding-box center coordinates and area, and class probabilities, giving dimension m = n + 4.Here n is the number of possible classes.Source: §3.5 · p. 10

3.6. Anomaly Detection

  • The detector learns nominal patterns from training features using kNN distances, with a percentile distance as a baseline for test-instance anomaly evidence.Training assumes anomaly-free data and partitions the nominal feature set to calculate distances; in testing, the resulting evidence updates a CUSUM-like statistic, with positive successive evidence expected to grow it.Source: §3.6 · pp. 11–12

3.7. Threshold Selection

  • The method derives an asymptotic upper bound on average false alarm rate and uses it to select a decision threshold for a desired false alarm period.The text says the bound's terms can be computed from training data and the feature dimensionality, enabling threshold selection from the bound.Source: §3.7 · pp. 12–13

4. Experiments

The experiments evaluate MONAD on three public datasets, compare it with existing methods, and examine its sequential decisions, online ROC performance, false-alarm bound, and computational cost.

4.1. Datasets

  • The evaluation uses CUHK Avenue, UCSD Ped2, and ShanghaiTech, treating each dataset independently because their characteristics and challenges differ.UCSD Ped1 is excluded; Avenue and Ped2 include 16 training videos each, while ShanghaiTech includes 330 training videos across 13 scenes.Source: §4.1 · pp. 13–14

4.3. Implementation Details

  • The implementation uses a U-NET generator, patch discriminator, and pretrained YOLOv3 object detector, with the sequential anomaly detector implemented in Python.Frames are resized to 256 × 256; detector confidence cutoffs are greater than 0.6 for training and at least 0.4 for testing.Source: §4.3 · pp. 14–15

4.4. Impact of Sequential Anomaly Detection

  • Instantaneous anomaly evidence is more prone to false alarms than MONAD’s sequential statistic, which integrates recent evidence to handle noise.The nonsequential comparator thresholds the instantaneous evidence δt.Source: §4.4 · p. 15, Figure 3 · p. 16

4.5. Results

  • On UCSD and ShanghaiTech, MONAD achieves higher true alarm rates than the online version of within practical false-alarm rates.The comparison uses ROC curves over the practical false-alarm range.Source: §4.5 · p. 16, Figure 4 · p. 18
  • MONAD outperforms existing results on Avenue and UCSD and achieves competitive performance on ShanghaiTech.Table 1 compares frame-level AuROC across the three datasets.Source: §4.2 · p. 14, §4.5 · p. 15, Table 1 · p. 17
  • Experiments corroborate the derived false-alarm bound and the procedure for selecting threshold h.Figure 5 presents actual false-alarm periods against lower bounds; the plotted quantity is the logarithm of the period, the inverse of false-alarm rate.Source: §4.5 · p. 17

4.6. Computational Complexity

  • The sequential detector’s training time complexity is O(M1M2m), while testing is O(M2m); the deep-learning frame-prediction module runs at 22 frames per second.The reported YOLO object-detector time is about 12 milliseconds per image, or about 83.33 frames per second; tests used an RTX 2070 and i7-8700k.Source: §4.6 · pp. 17–18

5. Conclusion

The conclusion presents MONAD as an online video-anomaly detector combining deep-learning feature extraction and statistical decision making. Its experiments show competitive benchmark performance and false-alarm rates consistent with the derived asymptotic bounds.

5. Conclusion

  • MONAD combines deep-learning feature extraction with statistical decision making and is presented as an online video-anomaly algorithm.Its feature extractor combines GAN-based frame prediction with a lightweight object detector, while its sequential detector supports performance analysis.Source: §5 · p. 18
  • Experiments show that MONAD outperforms state-of-the-art methods and yields false-alarm rates consistent with the derived asymptotic bounds.Figure 5 compares actual false-alarm periods with lower bounds for UCSD Ped2, ShanghaiTech, and Avenue.Source: §5 · pp. 18–19, Figure 5 · p. 19
  • Future work will address timely detection using a metric based on average delay and precision.Source: §5 · p. 19

Appendix A. Proof of Theorem 1

The proof analyzes nominal frame-level anomaly evidence under an independence assumption, derives its asymptotic distribution from a Poisson point process, and uses it to obtain a false-alarm-period bound.

Appendix A. Proof of Theorem 1

  • For CUSUM-like algorithms with independent increments, the proof uses a lower bound on the average false-alarm period.In the nominal case, it assumes δt is independent over time because there is no anomalous object to correlate selections.Source: Appendix A · p. 20
  • The proof derives the nominal distribution of frame-level evidence from nearest-neighbor distances in a Poisson point process.The maximum-distance selected instances are treated as independent over time; the nearest-neighbor distribution uses the Poisson rate measure for a ball around F_t.Source: Appendix A · p. 20
  • The derivation uses the density of δt and the Lambert-W function to obtain ω0, which is then used with threshold h in the false-alarm-period bound.The training set supplies the upper bound φ, and the proof states that the false-alarm rate is the inverse of the false-alarm period.Source: Appendix A · pp. 20–22

Related papers

Loading 2010.07110v1…