Online Detection of Abnormal Events Using Incremental Coding Length

Jayanta Kumar Dutta, Bonny BanerjeeAAAI 2015 · DOI

Abstract

This paper presents an unsupervised, online framework for detecting anomalies in streaming video by measuring the "rarity" of spatiotemporal features. By integrating sparse dictionary learning with Incremental Coding Length (ICL), the system identifies abnormal events—such as unauthorized vehicles on walkways or sudden crowd dispersals—without requiring labeled training data or extensive video history. The method adapts to non-stationary distributions in real-time, achieving state-of-the-art performance on standard benchmarks.

Introduction

Detecting abnormal events in surveillance video is challenging due to the inherent rarity of anomalies, the massive volume of data, and "concept drift," where the definition of "normal" changes over time. Traditional supervised methods are often impractical because they require large, labeled datasets of rare events. This work proposes a system that learns a model of normality incrementally and detects outliers in real-time without human intervention, making it ideal for smart sensors with limited storage.

Methodology

The proposed pipeline is based on sparse coding and information theory. The system first extracts "spatiotemporal cuboids" (small 3D chunks of video) around detected interest points to capture both appearance and motion.

Spatiotemporal interest point detection across various benchmark datasets including UMN, UCSD, and Subway.

Instead of a fixed dictionary, the system employs Online Sparse Dictionary Learning to evolve as new video frames arrive. The core innovation is the use of Incremental Coding Length (ICL) to score abnormalities. An event is flagged if it consists of features that are rarely used (measured by entropy gain via ICL) but are currently appearing with high reconstruction intensity. This joint consideration of rarity and strength allows the system to avoid manual parameter tuning.

Experiments and Results

The researchers evaluated their approach on three standard benchmarks: the UCSD Pedestrian dataset, the UMN crowd dataset, and the Subway entrance/exit dataset.

  • UMN Dataset: The system achieved an Area Under the Curve (AUC) of 99.5% and an Equal Error Rate (EER) of 3.65%, performing on par with state-of-the-art (SOTA) methods in detecting sudden crowd dispersals.
  • UCSD Ped1: Recorded an EER of 19.8% and a Rate of Detection (RD) of 69.5% for anomalies like bikers and skaters on walkways.
  • Subway Dataset: Successfully detected nearly all "wrong direction" and "no payment" events (e.g., 19/19 events in the Exit gate video) with very few false alarms.
The system successfully localizing and flagging abnormal crowd activity in the UMN dataset.

Discussion and Conclusion

A key insight of this work is that rarity alone is insufficient for anomaly detection; it must be combined with significant contribution to the current scene's reconstruction. The system is memory-efficient, requiring only small summary matrices rather than raw video history. While the method's performance is currently tied to the quality of interest point detection and fixed cuboid scales, its unsupervised and online nature makes it a robust candidate for real-world streaming applications where data distributions shift over time.

Summary updated

BibTeX
@inproceedings{dutta2015online,
  author = {Jayanta Kumar Dutta and Bonny Banerjee},
  title = {{Online Detection of Abnormal Events Using Incremental Coding Length}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9799},
}
Figure