Object Detection at Level Crossing Using Deep Learning.

Area of Science:

• Railway Engineering
• Artificial Intelligence
• Computer Vision

Background:

• Rail industry expansion increases level crossing risks, leading to accidents involving pedestrians and vehicles.
• Traditional sensors at level crossings are limited, often requiring multiple units and using ineffective algorithms for dynamic environments.
• Existing safety systems struggle with unpredictable user behavior, contributing to fatalities and near misses.

Purpose of the Study:

• To address the safety risks at rail level crossings by proposing an advanced sensing system.
• To integrate deep learning technology with existing vision systems for enhanced object detection and localization.
• To evaluate the effectiveness of a novel sensing system in a complex and dynamic environment.

Main Methods:

• Integration of deep learning models (e.g., MobileNet) with CCTV systems for object detection and localization.
• Development and discussion of algorithms for post-processing sensor information.
• Exploration of radar systems for fail-safe interlocking mechanisms.

Main Results:

• The proposed deep learning model achieved approximately 88% accuracy in classification using MobileNet.
• The object detection component of the model yielded a loss metric of 0.092.
• The system demonstrated the capability to detect and localize various objects like pedestrians, bicycles, and vehicles.

Conclusions:

• Deep learning integration significantly enhances object detection capabilities at rail level crossings.
• The proposed system offers a promising solution for improving safety and reducing incidents.
• Further research and development can build upon these findings for more robust safety solutions.