Enhanced Few-Shot Learning for Intrusion Detection in Railway Video Surveillance

Xiao Gong, Xi Chen, Zhangdui Zhong, Wei Chen

TL;DR

This paper tackles the challenge of railway intrusion detection in video surveillance, where limited intrusion examples and highly similar scenes make traditional deep learning difficult. They propose an enhanced few-shot learning approach using Model-Agnostic Meta-Learning (MAML) combined with novel techniques like mask inputs, label shuffling, and neuron dropout to achieve robust detection even with very few training samples.

What Problem Does This Paper Address?

Railway intrusion detection is crucial for safety, but current video surveillance systems face significant hurdles. The primary challenge is the scarcity of "intrusion" data (e.g., people on tracks) compared to normal "non-intrusion" footage. This imbalance, coupled with the inherent similarity of railway scenes (e.g., the same track layout captured by a single camera) and the diversity of different railway environments, makes it difficult for standard deep learning models to generalize. These models typically require vast amounts of labeled data to learn effectively, which is impractical and costly to collect for rare intrusion events and diverse railway landscapes.

Methods

The core of the proposed solution is an enhanced few-shot learning framework built upon the Model-Agnostic Meta-Learning (MAML) algorithm. MAML is designed to learn how to learn, enabling models to adapt quickly to new tasks with minimal data. The authors enhance MAML in several key ways:

  • Feature Engineering with Mask Inputs: They introduce a pre-processing step using a Fully Convolutional Network (FCN) to extract "masks" that highlight essential track areas. These masks are then concatenated with the original RGB video frames, providing the meta-learner with focused spatial information.
  • Mitigating Overfitting: To combat the high intra-scene similarity (leading to overfitting), they employ two techniques:
  • Shuffling Labels: Randomly shuffling intrusion/non-intrusion labels during meta-training prevents the model from memorizing specific class assignments.
  • Neuron Dropout: Applying dropout to the linear layer of the CNN helps regularize the model and improve generalization.
  • Modified CNN Architecture: The CNN architecture is adapted to handle the higher resolution of railway video frames (640x480) compared to standard benchmarks like MiniImagenet (84x84).

Datasets

  • Local Railway Scene Dataset: A custom dataset collected from PTZ cameras at 65 railway "throat areas" in China, spanning December 2018 to April 2019. This dataset is characterized by low inter-scene dissimilarity and high intra-scene similarity.
  • MiniImagenet Dataset: A standard benchmark dataset for few-shot learning, used for comparative analysis and to demonstrate the challenges posed by the railway dataset's characteristics.

Key Insights and Ideas

  • Few-shot learning is a viable solution for rare event detection: The paper demonstrates that meta-learning techniques can effectively address the data scarcity problem in railway intrusion detection.
  • High intra-scene similarity is a major overfitting culprit: The authors theoretically and empirically show that similar training samples lead to similar gradients, making it hard for meta-learners to adapt to new tasks.
  • Masks as focused spatial cues significantly improve detection: By explicitly highlighting track areas, the FCN-generated masks guide the meta-learner to focus on relevant features, improving accuracy.
  • Label shuffling and dropout are crucial for generalization: These techniques are shown to be effective in mitigating overfitting, a common issue in meta-learning with imbalanced or similar data.
  • MAML's adaptability is key, but needs enhancement for specific domains: While MAML provides a strong foundation, domain-specific challenges require tailored improvements.

Key Results

  • The proposed enhanced few-shot learning algorithm significantly outperforms standard randomly initialized supervised learning and classical meta-learning approaches (like ProtoNets) on the railway intrusion detection task.
  • The method achieves high accuracy even with very few training samples (e.g., K=1 to K=5). For instance, with K=5, the proposed method achieves 83.80% accuracy, outperforming ProtoNets (78.27% to 83.80% depending on settings).
  • The addition of mask inputs leads to a notable improvement in detection accuracy, reducing false positive rates.
  • The meta-learning approach with optimized initial conditions adapts much faster than random initialization, requiring fewer gradient steps to achieve comparable or better performance.

Limitations and Future Work

  • Computational Cost of FCN: While effective, the FCN segmentation step adds computational overhead, especially for real-time applications.
  • Generalization to Extreme Conditions: The paper notes that the FCN can struggle with challenging conditions not present in its training data, such as heavy snow or vertical shooting angles.
  • Future work could explore:
    • More efficient FCN architectures or alternative segmentation methods.
  • Robustness to a wider range of adverse weather and shooting conditions.
  • Further optimization of the meta-learning process for even faster adaptation.

Summary updated

Abstract

Video surveillance is gaining increasing popularity to assist in railway intrusion detection in recent years. However, efficient and accurate intrusion detection remains a challenging issue due to: (a) limited sample number: only small sample size (or portion) of intrusive video frames is available; (b) high inter-scene dissimilarity: various railway track area scenes are captured by cameras installed in different landforms; (c) high intra-scene similarity: the video frames captured by an individual camera share a same background. In this paper, an efficient few-shot learning solution is developed to address the above issues. In particular, an enhanced model-agnostic meta-learner is trained using both the original video frames and segmented masks of track area extracted from the video. Moreover, theoretical analysis and engineering solutions are provided to cope with the highly similar video frames in the meta-model training phase. The proposed method is tested on realistic railway video dataset. Numerical results show that the enhanced meta-learner successfully adapts unseen scene with only few newly collected video frame samples, and its intrusion detection accuracy outperforms that of the standard randomly initialised supervised learning.
BibTeX
@article{gong2021enhanced,
  author = {Xiao Gong and Xi Chen and Zhangdui Zhong and Wei Chen},
  title = {{Enhanced Few-Shot Learning for Intrusion Detection in Railway Video Surveillance}},
  journal = {IEEE Transactions on Intelligent Transportation Systems},
  year = {2021},
  doi = {10.1109/tits.2021.3102613},
}
Figure