Neural Network Training for Lane Boundary Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting traffic lane boundaries in autonomous vehicles are prone to errors due to low image resolution, weather conditions, and temporary occlusions, leading to unreliable detection and potential driving hazards.
Innovation Solution
A training method for a neural network that acquires and processes images from different orientations, simulates aerial views, and learns to differentiate between boundary elements and non-boundary elements, even in challenging conditions, using contextual information such as meteorological data and lighting conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional boundary detection algorithms are used, then the system is simple to implement, but detection reliability deteriorates under variable weather and lighting conditions
Solution Approach 1:
The patent applies preliminary action by collecting and storing boundary information from multiple sources (aerial images, road markings, signs) before the vehicle reaches the boundary area. This pre-collected information is stored in a database and used to predict boundary locations in advance, allowing the system to prepare reliable detection data before actual boundary detection is needed, thereby improving reliability without requiring complex real-time processing.
Solution Approach 2:
The patent transitions from two-dimensional image analysis to three-dimensional spatial reasoning by incorporating aerial images and geographic information system (GIS) data. This dimensional change allows the system to view boundary elements from multiple perspectives and combine them with vehicle position data, significantly improving detection reliability under various weather and lighting conditions.
2Measurement precision
If filtering modules are added to sort out bad detections, then detection accuracy improves, but the solution is curative rather than preventive and may be too late for safe driving decisions
Solution Approach 1:
The system performs preliminary action by collecting boundary information from multiple sources and storing it in a database before the vehicle reaches the boundary area. This pre-collected information includes aerial images, road markings, and signs, allowing the system to predict boundary locations in advance and provide accurate detection results before the vehicle needs to make driving decisions.
Solution Approach 2:
The patent implements feedback by continuously comparing predicted boundary locations with actual sensor data and updating the database with new information. This feedback loop ensures that the system learns from previous detections and improves its accuracy over time, providing both preventive and corrective mechanisms.
3Measurement precision
If multiple sensors are used to acquire images from different orientations, then detection precision improves, but device complexity increases
Solution Approach 1:
The patent applies universality by designing a integrated system where a single processing unit handles multiple data sources (aerial images, ground-based sensor images, GIS data). This multi-functional approach allows the system to process diverse information types through unified algorithms, improving boundary detection precision while managing complexity through consolidation rather than proliferation of separate processing systems.
Solution Approach 2:
The patent uses an intermediary approach by introducing a database as a mediator between multiple sensors and the processing algorithms. This database consolidates and standardizes data from various sources, allowing different sensors to work together seamlessly without requiring complex direct integration between each sensor and processing unit, thereby improving precision while managing system complexity.
Data Source
Figure 1
Figure 2
AI summary
This method trains a neural network that is required to detect boundaries of a lane of an automatically driven vehicle, and comprises: - a first step (10-20) in which first and second images of segments of a lane are acquired with a view to storage thereof, - a second step (30-70) in which, with some of the second images, an aerial image of the lane is constructed, then boundaries in this aerial image and elements in the first image are determined, then, if these elements extend these boundaries, they are stored in a database, each element otherwise being stored in this database as corresponding to the boundary, and - a third step (80) in which the neural network is trained with the records of the database.