RPIL-3D:3D localization for railway perimeter intrusion targets

Purpose This study aims to resolve the lack of three-dimensional (3D) spatial awareness in railway intrusion detection by proposing a monocular vision-based algorithm (RPIL-3D). It integrates geometric reconstruction (VGGT) and object detection (YOLOv8) to convert two-dimensional (2D) bounding boxes into geographic coordinates, enabling precise localization for risk assessment and early warning without relying on multisensor systems. Design/methodology/approach This paper proposes a 3D localization algorithm named RPIL-3D for railway perimeter intrusion targets, aiming to determine the spatial coordinates of intrusion targets in railway environments. The overall workflow consists of three key steps: First, multiple frames from railway surveillance video streams are input into the camera head module of the VGGT model to predict the camera’s intrinsic and extrinsic parameters from the monitoring images, thereby establishing the geometric relationship between the images and the 3D scene. Second, the YOLOv8 network is used to perform object detection on the surveillance footage, accurately identifying the 2D positions of intruders and generating bounding boxes. Finally, the pixel coordinates of the detection box center points are combined with the camera parameters, and through geometric projection and coordinate transformation formulas, they are mapped to actual latitude and longitude coordinates, achieving cross-domain localization from the image plane to real-world space. Findings This study addresses the challenge of lacking 3D spatial localization in railway intrusion scenarios by proposing RPIL-3D, a 3D localization algorithm for railway perimeter intrusion targets. The method achieves a fusion of 2D detection and 3D localization from images to geographic coordinates, enabling not only the identification of intrusion targets but also the determination of their precise spatial positions within railway environments. This capability provides critical support for subsequent risk assessment and hierarchical early warning based on spatial distance. Implemented under monocular vision conditions, the 3D localization functionality can be further extended to multicamera collaborative localization, integration with railway geographic information system (GIS) systems and coordination with intelligent dispatch systems. This significantly enhances the intelligence and practicality of railway intrusion detection systems. Originality/value Experimental results demonstrate that RPIL-3D effectively converts 2D detection results into 3D geographic coordinates, significantly improving the accuracy and practicality of intrusion localization. This provides reliable technical support for the intelligent upgrade of railway security monitoring systems.