Rail Detection via Particle Filtering Without Camera Calibration

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for detecting and tracking rails in video streams require camera calibration, are computationally intensive, and prone to errors due to variations in the ground plane and train movement, with complex setups and blurring issues affecting distant points, limiting their flexibility and reactivity.

Innovation Solution

A method using particle filtering to estimate weighted rail representation models directly in the image, eliminating the need for camera calibration, and modeling both rails and their spacing as a whole, with a cubic spline representation and linear function for distance calculation, allowing for real-time detection and tracking without homographic transformation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If homographic image transformation is used to obtain top-down view, then rail detection accuracy is improved, but computation time increases significantly and blurring occurs for distant points

Engineering Contradiction:
Improverail detection accuracyVSAvoidcomputation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts only the essential geometric relationships needed for rail detection (vanishing points, rail lines) without performing complete homographic transformation of the entire image. This selective extraction maintains detection accuracy while reducing computational burden by avoiding transformation of all image points.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

Instead of transforming the image to get top-down view and then detecting rails, the patent inverts the approach by directly detecting rail lines and vanishing points in the original perspective view, then using these detections to infer rail positions without full image transformation.

Inventive Principle:
Principle #13The other way round (Inversion)

2Measurement precision

If camera calibration is performed to eliminate ambiguity of furthest points, then detection precision is improved, but device complexity and instrumentation phase increase

Engineering Contradiction:
Improvedetection precisionVSAvoidinstrumentation phase
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs self-calibration by automatically detecting vanishing points from the image content itself rather than requiring external calibration equipment or procedures. The vanishing points are extracted directly from the rail line orientations in the image, making the system self-sufficient and eliminating complex instrumentation phases.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent combines multiple detection tasks (vanishing point detection, rail line detection, and calibration) into a unified process. The vanishing points detected from rail orientations serve dual purposes: they define the perspective geometry and simultaneously provide the calibration reference, merging what would otherwise be separate operations.

Inventive Principle:
Principle #5Merging (Combining)

3Stability of the object's composition

If Kalman filter is used for tracking, then tracking stability is improved, but reactivity to rail detection changes decreases and process complexity increases

Engineering Contradiction:
Improvetracking stabilityVSAvoidreactivity
Core Design Contradiction:
Stability of the object's compositionVSSpeed

Solution Approach 1:

The patent implements a dynamic tracking approach where the detection model adapts to changing rail configurations in real-time. Rather than using a fixed predictive model, the system continuously updates rail line parameters based on current frame detections, allowing rapid response to track changes while maintaining stability through consistent detection methodology.

Inventive Principle:
Principle #15Dynamics

4Device complexity

If rigid second-degree polynomial model is used for rail representation, then model simplicity is improved, but adaptability to various rail configurations decreases

Engineering Contradiction:
Improvemodel simplicityVSAvoidadaptability to rail configurations
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent segments the rail representation into multiple linear segments rather than using a single rigid polynomial model. Each segment can be independently fitted to local rail characteristics, allowing the model to adapt to curves, junctions, and straight sections while keeping individual segment computations simple and efficient.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3387619B1Detection and monitoring of rails in a video stream
Publication Date: 2020.01.08 COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES
  • EP3387619B1 patent drawingFigure 1
  • EP3387619B1 patent drawingFigure 2~3
  • EP3387619B1 patent drawingFigure 4~5

AI summary

A method for detecting and monitoring rails in images of a video stream filmed parallel to the axis of the rails, comprising, for at least one image of the video stream, the step of: Forming (E2) an image of rail contours (ICi), Characterised in that it comprises, for said at least one image of the video stream, the steps of: Estimating, by particle filtering (E3), a set of weighted models of representations of rails (Mj), as a function of the image of rail contours, each model being defined by parameters which comprise the x-coordinates of control points of a representation of a central axis of the rails in a frame of reference defined in the image, the y-coordinates of the control points being predefined and distinct, and coefficients of a linear function representing the distance between the representation of the central axis of the rails and, respectively, representations of the left and right rails as a function of the y-coordinate in the frame of reference of the image, and Estimating (E4) the position of the rails in the image as a function of the set of weighted models.