Synthetic Data Generation for Autonomous Driver Testing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for testing advanced driver-assistance systems (ADAS) require capturing real data from various sensor positions and characteristics, which is costly and time-consuming, especially when sensors' physical characteristics change or when producing large datasets, and existing synthetic data methods often result in distorted signals due to inadequate depth maps and alignment issues.
Innovation Solution
A system that generates synthetic data by capturing multiple signals from a common scene, creating dense depth maps, applying point of view transformations, and incorporating style and geometric distortions to simulate signals from a target sensor's perspective, reducing distortion and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real data is captured from various sensor positions and characteristics, then testing accuracy is improved, but cost and time increase
Solution Approach 1:
The patent creates synthetic copies of real sensor data through simulation. A physics-based simulation engine generates synthetic sensor signals that replicate the characteristics of real sensor data without requiring actual physical data collection. This allows testing to be performed on synthesized datasets that mimic real-world conditions, significantly reducing the time and cost of capturing real data while maintaining testing accuracy.
Solution Approach 2:
The simulation engine allows dynamic adjustment of sensor parameters such as position, orientation, field of view, and sensitivity characteristics. By changing these parameters in the simulation, the system can generate data for various sensor configurations without physically repositioning sensors or recapturing real data, thereby maintaining testing accuracy across different scenarios while minimizing time and resource expenditure.
2Reliability
If real data is captured from various sensor positions and characteristics, then testing accuracy is improved, but cost increases
Solution Approach 1:
The system replaces expensive real-world data collection with cost-effective synthetic data generation. By using a physics-based simulation engine to create realistic sensor data, the patent eliminates the need for costly physical data capture campaigns, sensor installations, and field testing while maintaining the fidelity required for accurate ADAS testing.
Solution Approach 2:
The simulation engine generates disposable synthetic datasets that can be created on-demand without the recurring costs associated with real data collection. Each synthetic dataset can be generated independently and discarded after use, eliminating the need for expensive, reusable physical testing infrastructure while maintaining testing accuracy.
3Productivity
If synthetic data is generated without adequate depth maps, then production speed is improved, but signal distortion increases
Solution Approach 1:
The simulation engine pre-computes accurate depth maps and spatial relationships as part of the synthetic data generation process. By establishing the three-dimensional scene geometry and sensor-to-scene transformations in advance, the system ensures that depth information is accurately embedded in the synthetic data from the outset, preventing signal distortion while maintaining efficient production speeds.
Solution Approach 2:
The patent introduces a physics-based simulation engine as an intermediary between the scene description and the generated sensor data. This intermediary layer computes accurate depth maps, occlusions, and geometric transformations, ensuring that the synthetic data maintains spatial accuracy and avoids distortion while being generated efficiently through algorithmic processes.
Data Source
AI summary
A system for creating synthetic data for testing an autonomous system, comprising at least one hardware processor adapted to execute a code for: producing a plurality of synthetic training signals, each simulating one of a plurality of signals simultaneously captured from a common training scene by a plurality of sensors, and a plurality of training depth maps each qualifying one of the plurality of synthetic training signals according to the common training scene; training a machine learning model based on the plurality of synthetic training signals and the plurality of training depth maps; using the machine learning model to compute a plurality of computed depth maps based on a plurality of real signals, the plurality of real signals are captured simultaneously from a common physical scene, each of the plurality of real signals are captured by one of the plurality of sensors, each of the plurality of computed depth maps.