Neural Network Validating Vehicle Environment Map
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated driving systems face challenges in accurately and reliably validating environment maps using sensor data, which is crucial for safe operation, especially due to the difference in representation between map data and sensor data.
Innovation Solution
A method and system utilizing deep learning to train a neural network that directly validates a predefined environment map by converting map and sensor data into a joint latent space using convolutional neural networks, allowing for the comparison and validation of surroundings data without requiring an initial environment model, and determining validity through binary classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If map data and sensor data are used in different representation formats, then data processing is simpler, but validation accuracy deteriorates
Solution Approach 1:
The patent introduces a joint latent space as an intermediary representation that both map data and sensor data are transformed into. This latent space serves as a common language that preserves the semantic meaning of both data types while enabling direct comparison. The neural network learns this shared representation space during training, allowing accurate validation without requiring the original data to be in the same format.
2Reliability
If an environment model is created from sensor data first, then validation can be performed, but processing time increases
Solution Approach 1:
The patent performs preliminary action by training the neural network offline with pre-collected map and sensor data pairs. During this training phase, the network learns to directly compare map representations with sensor observations and predict their compatibility. This pre-training eliminates the need for complex environment model creation during actual validation operations, enabling fast real-time validation.
Solution Approach 2:
The patent replaces the traditional mechanical approach of creating explicit environment models with a data-driven neural network approach. Instead of manually or algorithmically constructing environment models from sensor data, the system uses a trained neural network to directly predict map-sensor compatibility, significantly reducing processing time while maintaining validation reliability.
3Measurement precision
If deep learning is used to train the neural network, then validation accuracy improves, but computational complexity increases
Solution Approach 1:
The patent transforms the validation problem from comparing raw high-dimensional map and sensor data to comparing their lower-dimensional latent space representations. By changing the parameter space from original data dimensions to compressed latent features, the system achieves high validation accuracy with reduced computational complexity. The neural network learns optimal parameter transformations during training.
Data Source
AI summary
A method is described for providing a neural network for directly validating an environment map in a vehicle by means of sensor data. Valid or legitimate environment data is provided in a feature representation from map data and sensor data. Invalid or illegitimate environment data is provided in a feature representation from map data and sensor data. A neutral network is trained using the valid environment data and the invalid environment data.