Research on train anti-collision method based on deep learning

Jianming Zhang, Qijin Lu, Shuting Zao2018 · DOI

Abstract

Real-time detection of train distance from the front of the obstacle to ensure that the train in the case of safe braking speed driving is to improve the efficiency of train driving the main way.

Introduction

In modern rail transit, safety is typically managed by complex ground-based signaling systems like Automatic Train Protection (ATP). However, when these systems fail or when trains operate in manual or emergency modes, drivers lose their primary safety net. Traditional positioning methods—such as speed sensors, radar, and track circuits—can be expensive to install as backups or may lack the "intelligence" to recognize specific obstacles and signal colors visually.

This paper proposes a deep-learning-based computer vision system designed as a fail-safe backup for train anti-collision and signal monitoring. By utilizing a dual-camera setup (long and short focus) and a calibrated distance-measurement algorithm, the system provides real-time obstacle detection and speed recommendations, ensuring safety even during primary signaling system failures.

Methodology

The proposed system relies on a multi-stage image processing pipeline and a hybrid distance-measurement approach.

Image Acquisition and Classification

The system uses both long-focus and short-focus cameras to capture a wide range of the environment in front of the train. Deep learning classifiers were established for four key categories: tracks, trains, turnouts, and signals. The pipeline includes standard preprocessing (grayscale conversion, edge detection, and noise removal) followed by feature extraction to identify these objects within video frames.

Detection of railway tracks and turnouts using deep learning

Distance Measurement and Calibration

A hybrid approach is used for distance estimation:

  1. Initial Calibration: A laser rangefinder measures the precise distance to specific points on the track and other trains.
  2. Pixel-to-Distance Mapping: These measurements calibrate the system, mapping specific pixel rows in the camera image to physical distances.
  3. Real-time Estimation: Once calibrated, the system estimates the distance to an obstacle based on its pixel position in the image, removing the need for continuous laser firing.
Real-time track identification in a subway environment

Decision Logic

The system monitors signal colors and obstacle proximity. If a red light is detected or an obstacle is within a critical range, it calculates a recommended speed based on maximum braking distances and triggers voice alarms for the driver.

Results and Discussion

The system was validated through field tests on an actual subway lane. Key findings include:

  • Functional Accuracy: The method accurately determined obstacle distances, orbit environments, and signal statuses across multiple test points.
  • Safety Improvements: The system successfully provided voice prompts and alarms, improving the collision avoidance probability (referred to as the "train defense hit coefficient").
  • Operational Efficiency: By providing recommended speeds and routes during signal failures, the system reduces reliance on manual supervisors and improves recovery time.

The most significant contribution is the move toward "independent control" layers. By making the train's safety logic independent of trackside infrastructure, the system becomes a robust fallback during network or signal failures.

Limitations and Future Work

While the field tests were successful, the paper lacks specific quantitative metrics like Mean Average Precision (mAP) or exact distance error margins. Future work should address environmental robustness (e.g., handling dust, smoke, or glare) and scale the testing to broader road networks to ensure long-term reliability.

Summary updated

Abstract

Real-time detection of train distance from the front of the obstacle to ensure that the train in the case of safe braking speed driving is to improve the efficiency of train driving the main way.
BibTeX
@misc{zhang2018research,
  author = {Jianming Zhang and Qijin Lu and Shuting Zao},
  title = {{Research on train anti-collision method based on deep learning}},
  year = {2018},
  doi = {10.2991/iceesd-18.2018.198},
}
Figure