Medium-Fidelity Vision Sensor Simulation for ADAS
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for designing and testing Advanced Driver Assistance Systems (ADAS) and autonomous vehicle systems are hindered by the expense and limitations of using real-world image data, which cannot be easily modified to simulate variations in scenes, and existing simulation methods, such as point target models and high-fidelity gaming engine models, are either low fidelity or computationally expensive.
Innovation Solution
A system and method for generating simulated object detection data by modeling a vision sensor and detection algorithm, using adjustable parameters to create a medium-fidelity model that accounts for sensor intrinsic and extrinsic parameters, occlusion, and measurement noise, allowing for efficient simulation of object detection in various scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-fidelity gaming engine models are used for simulation, then measurement precision and realism are improved, but computational cost and processing time increase significantly
Solution Approach 1:
The patent applies local quality by selectively modeling only the critical aspects of vision sensor detection at medium fidelity rather than comprehensively simulating all physical phenomena. The model focuses on geometric relationships, occlusion detection, and basic sensor parameters while omitting computationally intensive elements like realistic lighting, materials, and complex environmental effects, thereby achieving adequate detection accuracy at reduced computational cost.
Solution Approach 2:
The patent implements dynamics by making the simulation model adjustable and configurable. Parameters such as sensor field of view, detection ranges, occlusion thresholds, and measurement noise characteristics can be dynamically modified to match different sensor configurations and test scenarios, allowing the model to adapt its fidelity level based on computational resource availability and test requirements.
2Measurement precision
If real-world image data is used for testing, then measurement precision is improved, but adaptability and ease of modification decrease
Solution Approach 1:
The patent creates simplified copies of real-world scenes using geometric primitives (planes, boxes, cylinders) that replicate essential spatial relationships and occlusion patterns without requiring actual image data. These synthetic scene representations can be programmatically generated and modified, providing both the accuracy needed for detection algorithm testing and the adaptability to create varied test scenarios through code rather than manual image selection.
Solution Approach 2:
The patent enables scene variation through parameter changes by allowing programmatic adjustment of geometric object positions, dimensions, orientations, and material properties. Test scenarios can be systematically varied by modifying numerical parameters such as object coordinates, sensor mounting positions, field of view angles, and environmental conditions, facilitating comprehensive algorithm evaluation across diverse conditions without requiring new real-world image datasets.
3Productivity
If low-fidelity point target models are used, then computational efficiency is improved, but measurement precision and reliability deteriorate
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing geometric relationships, occlusion matrices, and sensor parameter configurations before actual detection algorithm testing. The medium-fidelity model prepares detailed scene representations with accurate geometric data and occlusion information in advance, allowing faster processing during the actual test execution while maintaining higher precision than point target models that lack these pre-computed details.
Solution Approach 2:
The patent creates a composite simulation model that combines elements of both low-fidelity and high-fidelity approaches. It integrates the computational efficiency of simplified geometric representations with enhanced features for occlusion detection, sensor parameter modeling, and measurement noise simulation. This composite model achieves better detection accuracy than pure point target models while maintaining faster processing than complete high-fidelity gaming engine simulations.
Data Source
AI summary
Synthetic object detection data is generated for a modeled sensor, such as a camera. Scenario data specifying objects, such as vehicles, sensor intrinsics, such as focal length, principal point, and image size, and sensor extrinsics, such location and orientation in the scenario of the sensor, may be received. An object detector model may detect a given object in the scenario if it lies within the sensor's field of view, is large enough, and is not occluded. Two dimensional (2D) image plane position and velocity measurements may be generated. A measurement noise model may add noise to the measurements. Position, velocity, and measurement noise may be mapped into a three dimensional (3D) world coordinate system. An object detection list that includes time of detection, detected position and velocity, measurement accuracy, and an object classification for detected objects may be output.