Autonomous Vehicle ROI Extraction Using 3D Drivable Region Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous vehicle systems using high-resolution cameras face challenges in maintaining inference speed while preserving image quality, often requiring downsizing of images, which degrades environmental information, and fixed region-of-interest methods fail to adapt to changing driving scenarios.
Innovation Solution
A method and system that extract a region of interest based on 3D drivable region information from high-resolution cameras, using far-distance point groups, image projection, dimension reduction, weighting calibration, and region-of-interest search algorithms to identify the most relevant drivable areas for processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution camera is used to acquire detailed information, then image quality and environmental information are improved, but inference time increases substantially
Solution Approach 1:
The patent extracts only the necessary region of interest from the high-resolution image for neural network inference, rather than processing the entire image. This is achieved by identifying the drivable region and selecting the farthest point as the center of the ROI, then extracting a cropped region around this center point. This extraction approach maintains image quality for the relevant area while significantly reducing the computational load and inference time.
Solution Approach 2:
The patent segments the high-resolution image into multiple regions, identifying and processing only the region of interest that contains the farthest drivable area. By dividing the image processing task into segments (full image analysis for ROI identification, then focused processing of only the extracted ROI), the system achieves both high measurement precision in the critical region and reduced overall processing time.
2Productivity
If image downsizing is applied to guarantee inference speed, then processing speed is improved, but image quality and environmental information are degraded
Solution Approach 1:
Instead of downsizing the entire high-resolution image, the patent extracts only the necessary region of interest at full resolution for neural network inference. This extraction-based approach selectively reduces the amount of data processed while preserving the full image quality and environmental information in the critical drivable region, avoiding the quality degradation that would result from blanket downsizing.
3Ease of operation
If fixed region of interest is used, then processing simplicity is improved, but adaptability to changing driving scenarios deteriorates
Solution Approach 1:
The patent implements a dynamic region of interest extraction method where the ROI center is determined by identifying the farthest point in the drivable region, which changes based on the current driving scenario, road geometry, and environmental conditions. This dynamic approach allows the system to automatically adapt to curved lanes, intersections, and other varying driving situations while maintaining a relatively simple processing framework based on automated farthest-point identification.
Data Source
AI summary
There are a method and a system for extracting a region of interest in an image obtained through a high-resolution camera installed in an autonomous vehicle. The method for extracting the region of interest based on a drivable region in the high-resolution camera includes a step of acquiring an image through the high-resolution camera; and a step of extracting, by a processor, a region of interest within the acquired image by using 3D drivable region information. Accordingly, far-distance information extraction performance may be enhanced, and all of the functions of a set of a wide angle camera and a narrow angle camera of HD may be performed only with one FHD or UHD wide angle camera, so that the number of sensors may be reduced and a production cost may be reduced.