Consistent Sparse Representation for Abnormal Event Detection

Zhong Zhang, Shuang Liu, Zhiwei ZhangIEICE Trans. Information and Systems 2015 · DOI

Abstract

Abnormal event detection in video surveillance is a critical yet challenging task due to the rarity of anomalies and the high dimensionality of video data. This paper introduces Consistent Sparse Representation (CSR), a novel framework that addresses these challenges through a two-pronged approach: neighborhood-preserving dimensionality reduction and a consistency-penalized sparse coding objective. By forcing normal event representations to cluster around a collective mean, CSR creates a more distinct boundary between normal and abnormal behaviors. Evaluations on the UCSD Ped1 and Subway datasets show that CSR achieves state-of-the-art performance, significantly improving detection rates while minimizing false alarms.

Introduction

Detecting anomalies like unauthorized vehicles on pedestrian paths or fare-dodging in subways typically relies on "one-class" learning, where a model is trained exclusively on normal data. However, traditional sparse representation methods often struggle with the "curse of dimensionality" and a lack of relationship modeling among training samples. Standard sparse coding treats each sample independently, often resulting in a "blurry" boundary that leads to high false-positive rates or missed detections.

Methodology

The CSR framework proposes two primary innovations to refine the representation of normal events and improve anomaly discriminability.

1. Neighborhood-Preserving Dimensionality Reduction

To handle high-dimensional video features, the authors first partition training data into $k$ clusters using K-means. Each feature vector is then projected into a lower-dimensional space by calculating normalized distances to these cluster centers via a Gaussian kernel. This ensures that features proximate in the original space remain close in the reduced space, maintaining local structures more effectively than global methods or standard Locally Linear Embedding (LLE).

2. Consistent Regularization

The core of the CSR method is a modified sparse learning objective. In addition to reconstruction error and an $l_1$-norm sparsity penalty, the authors introduce a Consistent Regularization term. This term penalizes the distance between an individual sample's coefficient vector and the mean vector of all normal samples.

Conceptual illustration of consistent regularization: (a) standard sparse representation vs. (b) consistent sparse representation where normal samples (green) are pulled toward a common mean, enhancing the separability of anomalies (red stars).

This "pull" factor forces the model to learn a compact latent representation for normal events. By shrinking the variance of normal samples in the feature space, outliers (anomalies) become significantly more distinguishable, effectively increasing the margin for detection.

Experimental Results

The method was rigorously tested on two standard benchmarks: the UCSD Ped1 and Subway datasets.

On the UCSD Ped1 dataset, CSR achieved a frame-level Equal Error Rate (EER) of 13.3% and an Area Under Curve (AUC) of 93.1%. At the pixel level, the improvement was even more pronounced, with the Equal Detected Rate (EDR) reaching 60.2%, a substantial jump from the 48% achieved by previous sparse methods.

Quantitative comparison of EDR and AUC on the UCSD Ped1 dataset (pixel-level), highlighting the superior performance of the proposed CSR method.

In the Subway dataset, which involves complex behaviors like "wrong direction" or "no-pay" events, CSR detected 35 out of 12 total events across entrance and exit gates while maintaining the lowest false alarm rate (4/0) among all compared algorithms.

Conclusion

Consistent Sparse Representation offers a robust solution for anomaly detection by emphasizing the collective consistency of normal data. While the iterative optimization process and hyperparameter sensitivity ($\alpha$ and $\beta$) present opportunities for further refinement, the significant gains in detection accuracy and the reduction of false alarms mark a substantial advancement over traditional sparse coding techniques. Future work may focus on adapting the framework for real-time deployment and dynamic background environments.

Summary updated

BibTeX
@article{zhang2015consistent,
  author = {Zhong Zhang and Shuang Liu and Zhiwei Zhang},
  title = {{Consistent Sparse Representation for Abnormal Event Detection}},
  journal = {IEICE Trans. Information and Systems},
  year = {2015},
  doi = {10.1587/transinf.2015edl8113},
}
Figure