Obstacle Detection via Backprojection for Rail Vehicles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current obstacle detection methods for track-bound vehicles, especially rail vehicles, face limitations in reliability, complexity, and computational resource demands, particularly in automated driving scenarios, where they need to detect obstacles accurately and efficiently over long distances with varying environmental conditions.
Innovation Solution
A method and device for obstacle detection using back projection of image data based on reference 3D data, which reduces computational effort by determining back projection errors without the need for direct 3D data generation, leveraging known track profiles and relative poses of the image recording unit to identify obstacles in the clearance space.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If indirect obstacle detection using monocular camera technologies is employed, then device complexity and cost are reduced, but measurement precision and reliability deteriorate due to increased mathematical analysis requirements and environmental assumptions
Solution Approach 1:
The system pre-generates backprojection data from reference 3D data before obstacle detection. This preliminary computation creates a lookup structure that enables fast comparison with actual image data, eliminating the need for complex real-time 3D reconstruction while maintaining measurement precision.
Solution Approach 2:
The patent creates virtual depth information by backprojecting reference 3D data into image space. This copied geometric structure allows the system to compare expected pixel positions with actual obstacle pixels, achieving accurate distance measurement without direct 3D sensing hardware.
2Measurement precision
If 3D data generation methods are used for obstacle detection, then measurement precision improves, but computational effort and processing time increase significantly
Solution Approach 1:
Backprojection data is pre-computed from reference 3D data and stored for rapid access during detection. This shifts computational burden from real-time processing to offline preparation, enabling fast obstacle detection while maintaining 3D measurement precision.
Solution Approach 2:
The system computes backprojection data for the entire scene in advance, even though only portions containing obstacles need detailed analysis. This excessive pre-computation simplifies real-time detection by providing ready-to-use geometric references for all possible obstacle locations.
3Adaptability or versatility
If machine learning-based methods are applied for obstacle detection, then adaptability to different environments improves, but computational resources and training requirements increase heavily
Solution Approach 1:
The patent replaces machine learning-based obstacle detection with a geometric backprojection method. This deterministic approach uses physics-based light propagation geometry instead of data-driven models, eliminating training requirements and reducing computational resource consumption while maintaining environmental adaptability.
Solution Approach 2:
The backprojection method uses inherently available information (reference 3D data, camera pose, image data) to perform detection without external training data or complex model inference. The system serves itself by leveraging the geometric constraints already present in the sensing setup.
4Reliability
If direct obstacle detection using stereo camera systems or lidar is implemented, then measurement precision and reliability improve, but device complexity, cost, and energy consumption increase
Solution Approach 1:
The patent introduces backprojection data as an intermediary between the simple monocular camera and the desired 3D obstacle detection capability. This virtual depth map acts as a mediator that provides stereo-like measurement precision without requiring stereo cameras or lidar sensors.
Solution Approach 2:
The system replaces complex multi-sensor hardware systems (stereo cameras, lidar) with a computational approach using a single camera and pre-generated backprojection data. This substitutes mechanical sensing complexity with algorithmic processing while achieving equivalent measurement reliability.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
A method for obstacle detection for a lane-bound vehicle (1) is described. In the obstacle detection method according to the invention, first image data (BD1) from a light space (5) of the lane-bound vehicle (1) are acquired by an image acquisition unit (2) arranged on the lane-bound vehicle (1) with a first relative pose (RP1) of the image acquisition unit (2) relative to a lane (3) traversed by the lane-bound vehicle (1). Furthermore, second image data (BD2) are acquired by the image acquisition unit (2) with a second relative pose (RP2) of the image acquisition unit (2) relative to the lane (3) traversed by the lane-bound vehicle (1), which differs from the first relative pose (RP1).Furthermore, backprojection data (RPD), which assigns pixels of the first image data (BD1) and the second image data (BD2) to each other via backprojection, are determined based on previously known reference 3D data (PW) of the imaged light space (5). Finally, it is determined whether an obstacle (H) occurs in the light space (5), depending on a backprojection error (RPF), which is determined based on the pixels of the first image data (BD1) and the second image data (BD2) assigned to each other via backprojection. An obstacle detection device (50) is also described. A track-bound vehicle (1) is also described.