Intercity Railway Risk Space Anomaly Detection Based on Train Predeparture Key Frame Extraction and IADN Network
In contemporary intercity railway systems, the risk space between platform doors and trains is a critical safety concern due to potential hazards from abandoned items. Detecting anomalies within this space before train departure is vital for passenger safety and operational efficiency, highlighting the importance of precise and efficient methodologies despite the scarcity of abnormal samples.
Existing surveillance methods, primarily reliant on manual oversight, encounter notable challenges due to inherent blind spots and human limitations, resulting in increased false detection rates. Sensor-based approaches like infrared light curtains and laser detection systems have emerged but are susceptible to false alarms and blind spots. Similarly, vision sensor-based methods, while capable of real-time image processing, demand intricate algorithmic design and are labor-intensive. Deep neural networks provide a promising solution, but obtaining labeled anomaly data remains a formidable challenge.
In response to these challenges, the authors propose the Intercity Railway Risk Space Anomaly Detection (IRSA) framework based on Train Pre-departure which addresses these challenges by leveraging two core components of the framework: key frame extraction and an Image-inpainting Anomaly Detection Network (IADN).
Key frame extraction employs door position tracking and switching signal extraction to identify and extract crucial frames for anomaly detection. Operating within a critical 9-second timeframe, this process employs grayscale processing, binarization, and contour extraction facilitated by the Kernelized Correlation Filter (KCF) algorithm, ensuring precise determination of door positions.
The IADN network utilizes an unsupervised approach of completely erasing abnormal information to enhance detection accuracy. Complementing key frame extraction, the IADN network operates on an unsupervised basis, effectively eradicating abnormal information to heighten detection accuracy.
The network consists of several modules, including an image-inpainting Autoencoder (AE), a masking strategy, and a Global-attentive Reconstruction Error (GARE) module. IADN tackles anomaly detection challenges by directly using samples for training without labels, utilizing image-inpainting AE to extract standard information and fuse it with global reconstruction error for accurate detection.
The masking strategy divides images into patches, ensuring no duplicates, while the isotropic Vision Transformer (ViT) enhances processing efficiency. The decoder processes patches alongside mask tokens, computing loss, and structured similarity index (SSIM) loss to improve detection accuracy.
Experimental validation of the IRSA framework was conducted using the real-world IRSA dataset, comprising RGB videos sourced from intercity railway platforms.
The evaluation metrics included classification Area Under the Curve (AUC), localization precision, and recall, which demonstrated the framework's superior performance compared to existing methodologies. The framework achieved an overall classification AUC of 99.32%, localization precision of 97.45%, and recall of 96.29%, highlighting its efficacy in anomaly detection and localization.
The experimentation underscored the significance of factors such as visibility and obstructions in the train image, emphasizing their role in ensuring accurate anomaly detection. In conclusion, the Intercity Railway Risk Space Anomaly framework represents a significant advancement in anomaly detection technology, adeptly addressing longstanding challenges and ensuring safety and efficiency within intercity railway systems worldwide. Further advancements to the framework can find application in detection-based hazards in the motor, security, and locomotive industries.