Abstract
Railway track detection is a critical component for autonomous rail systems and safety monitoring. However, existing methods often struggle with the trade-off between high-speed processing and accuracy across varying camera perspectives. This paper introduces the Ultra-Fast Adaptive Track Detection Network (UFATD), a novel architecture that achieves extreme speeds (up to 473 FPS) by employing an adaptive row-selection approach. By classifying camera viewpoints and utilizing non-equidistant anchors, UFATD overcomes the rigidity of traditional models, providing robust real-time performance for high-speed trains and UAV-based monitoring.
Introduction
Autonomous railway operations and safety systems, such as those designed for trespasser detection, rely heavily on accurate track localization. Current research typically splits into two camps: segmentation-based models, which offer high accuracy but are computationally expensive, and anchor-based models, which are fast but often too rigid to handle perspective changes.
Standard anchor-based models utilize a fixed grid to identify tracks. When camera angles shift—a frequent occurrence with drone-mounted sensors or varying vehicle heights—the track may occupy only a fraction of the predefined anchors. This leads to wasted computation and poor detection of curved tracks near the horizon. UFATD addresses these challenges by creating a system that is both ultra-fast and perspective-adaptive.
Methodology
The UFATD architecture utilizes a backbone network (such as ResNet) feeding into two specialized branches designed for efficiency and adaptability.
Perspective Identifier
This branch functions as a classifier that analyzes the input image to determine the camera's viewpoint (e.g., high-angle vs. low-angle). Based on this classification, the model selects the most appropriate "anchor group" from a set of predefined options, ensuring the mathematical focus is placed where the track is most likely to be located.
Horizontal Coordinate Locator
Instead of dense pixel-wise segmentation, this branch performs row-based classification. It identifies the horizontal position of the track at specific vertical intervals (rows) defined by the chosen anchor group. This reduces a complex computer vision task into a series of simple classification problems.
Non-Equidistant Anchor Generation
A key innovation in UFATD is the use of non-equidistant anchors. Rather than spacing rows evenly, anchors are placed more densely near the horizon (the top of the image). This is where perspective causes tracks to appear to curve more sharply, allowing the model to capture fine details where they are most needed without increasing the overall computational load.
Experimental Setup and Datasets
The model was rigorously evaluated on three major datasets:
- SRail: A new dataset introduced by the authors containing 5,903 images, blending synthetic data from Unreal Engine 5 with real-world captures to simulate diverse camera angles.
- RailDB: A standard benchmark containing over 7,000 images.
- DL-RAIL: A benchmark dataset with 7,004 images.
Key Results and Insights
UFATD demonstrates that where a model looks is as critical as how it looks. By prioritizing the horizon and adapting to camera pitch, the network achieves state-of-the-art (SOTA) performance:
- Extreme Speed: The lightweight ResNet18-based version achieved 473 FPS, far surpassing real-time requirements.
- High Accuracy: The model reached an F1 score of 98.68% on the SRail dataset.
- Comparative Success: UFATD consistently outperformed leading models like UFLD and Rail-Net in both speed and accuracy across RailDB and DL-RAIL benchmarks.
- Ablation Benefits: The inclusion of the Perspective Identifier and the adaptive anchor algorithm provided a 3.5% boost to the mF1 score over the baseline.
Conclusion
UFATD provides a highly efficient solution for real-time railway monitoring. While the current model is optimized for rails, the adaptive anchor approach holds significant potential for other domains, such as lane detection in autonomous vehicles where road grades and camera angles vary. Future work will focus on enhancing feature fusion to improve performance in cluttered environments and further optimizing the anchor count to push processing speeds even higher.
Summary updated