Explainable Object Detection for Rail Wayside Object Verification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Safety-critical neural network object detection systems in rail vehicles struggle to provide explainable detections, making it difficult to identify misidentifications or non-detections of objects, which is a requirement for ensuring safety and reliability.
Innovation Solution
The integration of an explainable detection module within the object detection system, which uses classical computer vision methods to detect objects in a way that can be understood by humans, in conjunction with machine-learned detection using deep neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If neural network object detection is used, then detection accuracy and automation are improved, but explainability and human understandability deteriorate
Solution Approach 1:
The system segments the detection process into two independent modules: a neural network-based detection module that provides automated detection capabilities, and a classical computer vision module that generates human-understandable explanations. This segmentation allows each module to specialize in its strength while working together to resolve the contradiction between automation and explainability.
Solution Approach 2:
The patent introduces classical computer vision methods as an intermediary between the neural network detector and the human operator. This intermediary translates the opaque neural network outputs into interpretable visual evidence (edges, contours, shapes) that humans can understand and verify, thereby maintaining both automation and explainability.
2Productivity
If neural network object detection is used, then detection speed and productivity are improved, but safety and reliability deteriorate due to unexplainable detections
Solution Approach 1:
The system implements feedback by using classical computer vision检测结果 to verify and validate neural network detections. The classical module provides a second opinion on detected objects, creating a feedback loop that enhances reliability while maintaining the speed benefits of neural network processing.
Solution Approach 2:
The patent applies partial action by using classical computer vision methods selectively to verify critical detections rather than processing all detections equally. This approach maintains high detection speed while adding reliability checks where they are most needed for safety-critical applications.
3Device complexity
If only machine-learned detection is used, then device complexity is reduced, but the ability to identify misidentifications and non-detections deteriorates
Solution Approach 1:
The patent merges two different detection approaches (neural network and classical computer vision) into a unified system. This combination allows the system to leverage the strengths of both methods: the neural network's high-level pattern recognition and the classical methods' interpretable feature detection, thereby improving verification capability without excessive complexity.
Solution Approach 2:
The system achieves multi-functionality by designing a dual-module architecture where each module serves multiple purposes: the neural network provides fast detection while the classical module provides both verification and explanation generation. This universal design allows a single system to handle both automated detection and human-understandable verification.
Data Source
AI summary
A method of detecting a way-side object includes receiving data frames from sensors mounted on a vehicle. Position data corresponding to the position of the vehicle is received. Object-of-interest data is retrieved from a database. The sensor data frames and the object-of-interest data are processed to determine a region-of-interest in the sensor data frame. A portion of the sensor data frames corresponding to the region of interest is processed using machine-learned object detection to identify a first object-of-interest. The portion of the sensor data frames corresponding to the region of interest is processed using computer vision methods to detect features of the expected object-of-interest and identifying the detected object-of-interest as explained when the features of the expected object-of-interest are detected. Explained object-of-interest data corresponding to the explained detected object-of-interest is output to a navigation system of the vehicle.