Rolling-Shutter Point-Cloud Simulation With Zoned Capture Masks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for generating synthetic data for training perception systems in autonomous driving, such as ray-tracing, are resource-intensive and often fail to accurately simulate the rolling shutter effect, leading to poor quality training data due to high simulation rates that are beyond available computing resources.
Innovation Solution
The approach involves computing a sequence of partial simulation images, each associated with an estimated simulation time, and applying capture masks to simulate sensor point-clouds captured in a capture interval, reducing computational resources while increasing accuracy by dividing the scanning pattern into zones and using graphical processing units for image processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the simulation rate is increased to accurately simulate the rolling shutter effect, then the accuracy of simulated sensor point-clouds is improved, but the computational resources required increase beyond available capacity
Solution Approach 1:
The patent divides the scanning pattern into multiple zones and generates partial simulation images for each zone separately. This segmentation allows the system to process smaller regions with higher effective resolution without requiring a globally increased simulation rate, thereby maintaining accuracy while reducing overall computational burden.
Solution Approach 2:
The patent introduces a spatial dimension by dividing the field of view into zones and processing them independently. This dimensional approach allows selective refinement of specific regions rather than uniformly increasing simulation rate across the entire scene, achieving local accuracy improvements without proportional resource increases.
2Reliability
If the simulation rate is increased to capture rolling shutter effects, then the quality of training data is improved, but the cost and resource requirements become prohibitive
Solution Approach 1:
The patent applies different processing qualities to different zones. By identifying zones that contain relevant objects for training and applying higher-resolution partial simulation images specifically to those regions, the system improves training data quality where it matters most while using lower resources in less critical areas.
Solution Approach 2:
The patent generates partial simulation images with higher detail than a standard simulation rate would provide, but only for specific zones and time intervals relevant to the training objectives. This partial action approach delivers sufficient quality for reliable training without the excessive computational cost of uniformly high-rate simulation.
3Manufacturing precision
If full simulation images are generated at high simulation rates, then complete scene accuracy is improved, but the processing time and computational load increase significantly
Solution Approach 1:
The patent segments the scene into multiple zones and processes them as separate partial simulation images. This segmentation enables parallel processing of zones and allows the system to focus computational effort on specific regions, reducing total processing time while maintaining scene accuracy through comprehensive zone coverage.
Solution Approach 2:
The patent processes zones in periodic intervals rather than generating all simulation images simultaneously at high rate. By sequentially processing different zones at their required resolutions and combining them, the system achieves complete scene accuracy with distributed computational load over time, reducing peak processing requirements.
Data Source
AI summary
A system for generating synthetic data, comprising at least one processing circuitry adapted for: computing a sequence of partial simulation images, where each of the sequence of partial simulation images is associated with an estimated simulation time and with part of a simulated environment at the respective estimated simulation time thereof; computing at least one simulated point-cloud, each simulating a point-cloud captured in a capture interval by a sensor operated in a scanning pattern from an environment equivalent to a simulated environment, by applying to each partial simulation image of the sequence of partial simulation images a capture mask computed according to the scanning pattern and a relation between the capture interval and an estimated simulation time associated with the partial simulation image; and providing the at least one simulated point-cloud to a training engine to train a perception system comprising the sensor.