Online Least Squares One-Class Support Vector Machines-Based Abnormal Visual Event Detection

Tian Wang, Jie Chen, Yi Zhou, Hichem SnoussiSensors 2013 · DOI

TL;DR

This paper introduces novel online algorithms for detecting abnormal events in video surveillance. By combining "online least squares one-class support vector machines" (LS-OC-SVM) with a sparse representation technique, the method efficiently learns normal behavior and flags deviations in real-time, outperforming traditional batch learning approaches.

What Problem Does This Paper Address?

Real-time visual surveillance systems are crucial for safety and convenience, but automatically identifying "abnormal" events—behaviors that deviate from the norm—is a significant challenge. Traditional methods often rely on complex statistical models that can be difficult to obtain or are computationally expensive, especially when dealing with continuously arriving video data. The core problem is to develop a system that can learn what "normal" looks like from a stream of data and then quickly and accurately identify when something unusual happens, without requiring complete retraining for every new piece of information.

Methods

The paper proposes two main algorithmic contributions:

  • Online Least Squares One-Class Support Vector Machine (Online LS-OC-SVM): This method adapts the LS-OC-SVM framework for online learning. Instead of retraining the entire model with new data, it incrementally updates the model. This is achieved by leveraging the least squares formulation, which allows for solving a linear system instead of a quadratic programming problem, making updates more computationally efficient.
  • Sparse Online LS-OC-SVM: To further improve efficiency and handle large datasets, this approach introduces sparsity. It uses a "dictionary" of representative "normal" data points and approximates the model using these dictionary elements. A "coherence criterion" is employed to decide whether a new data point should be added to the dictionary, effectively controlling model complexity and preventing it from growing indefinitely.

Both methods use a covariance matrix descriptor to represent the movement information within each video frame, derived from optical flow. This descriptor captures the spatial arrangement and change rate of objects in the scene.

Datasets

The proposed methods were evaluated on:

  • Two-dimensional synthetic datasets: "Square" and "ring-line-square" distributions were used to test the algorithms' ability to learn complex shapes and detect deviations.
  • University of Minnesota (UMN) dataset: This is a benchmark dataset for abnormal crowd behavior detection, consisting of eleven sequences of crowded panic escape events recorded in lawn, indoor, and plaza scenes.

Key Insights and Ideas

  • Online learning is essential for real-time surveillance: The paper highlights the impracticality of retraining models from scratch with every new frame in video surveillance. Online learning offers a computationally feasible solution.
  • Least Squares SVM for efficiency: Shifting from traditional SVM's quadratic programming to LS-SVM's linear system for updates significantly speeds up the learning process.
  • Sparsity for scalability: The sparse online approach, by maintaining a dictionary of representative "normal" data, effectively manages model complexity and memory requirements, making it suitable for very large datasets.
  • Covariance descriptor as a robust motion feature: The covariance matrix descriptor, derived from optical flow, proves effective in capturing frame-level movement patterns for distinguishing normal from abnormal events.

Key Results

  • Competitive performance: Both the online LS-OC-SVM and sparse online LS-OC-SVM methods achieved high Area Under the ROC Curve (AUC) values on the UMN dataset, comparable to or exceeding state-of-the-art methods.
  • Lawn scene: The online LS-OC-SVM achieved an AUC of 0.9874, while the sparse online version reached 0.9510.
  • Indoor scene: Online LS-OC-SVM achieved an AUC of 0.8904, and sparse online LS-OC-SVM achieved 0.8886.
  • Plaza scene: Online LS-OC-SVM achieved an AUC of 0.9839, and sparse online LS-OC-SVM achieved 0.9515.
  • Efficiency: The sparse online strategy is noted for its computational efficiency and ability to adapt to memory constraints, making it suitable for handling large volumes of training data sequentially.

Limitations and Future Work

  • Distinguishing similar normal/abnormal states: The paper notes a limitation in distinguishing between abnormal sequences with no people and normal training samples that also have no people in certain image regions (e.g., upper half). This can lead to misclassification. The authors suggest that incorporating foreground information could help resolve this.
  • Computational cost of covariance descriptor: While effective, the computation of the optical flow and subsequent covariance descriptor can still be time-consuming. The paper mentions that using integral images can speed up covariance computation.
  • Comparison with SRC: While the proposed sparse online method is efficient, the paper acknowledges that Sparse Reconstruction Cost (SRC) [49] achieved higher AUC values on some metrics. However, SRC's approach of saving the entire training dataset in memory is not scalable. Future work could explore hybrid approaches that combine the efficiency of online learning with the accuracy of methods like SRC.

Summary updated

Abstract

The abnormal event detection problem is an important subject in real-time video surveillance. In this paper, we propose a novel online one-class classification algorithm, online least squares one-class support vector machine (online LS-OC-SVM), combined with its sparsified version (sparse online LS-OC-SVM). LS-OC-SVM extracts a hyperplane as an optimal description of training objects in a regularized least squares sense. The online LS-OC-SVM learns a training set with a limited number of samples to provide a basic normal model, then updates the model through remaining data. In the sparse online scheme, the model complexity is controlled by the coherence criterion. The online LS-OC-SVM is adopted to handle the abnormal event detection problem. Each frame of the video is characterized by the covariance matrix descriptor encoding the moving information, then is classified into a normal or an abnormal frame. Experiments are conducted, on a two-dimensional synthetic distribution dataset and a benchmark video surveillance dataset, to demonstrate the promising results of the proposed online LS-OC-SVM method.
BibTeX
@inproceedings{wang2013online,
  author = {Tian Wang and Jie Chen and Yi Zhou and Hichem Snoussi},
  title = {{Online Least Squares One-Class Support Vector Machines-Based Abnormal Visual Event Detection}},
  booktitle = {Sensors},
  year = {2013},
  doi = {10.3390/s131217130},
}
Figure