ROI-Based Object Recognition for Faster Autonomous Driving Inference

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current object recognition systems for autonomous driving face inefficiencies in processing large images, leading to prolonged inference times due to the high number of object candidate regions, which hinders real-time navigation and control of autonomous vehicles.

Innovation Solution

The implementation of a faster Region-based Convolutional Neural Network (R-CNN) with a Region Proposal Network (RPN) and a processor that extracts a road region of interest (ROI) to reduce the number of object candidate regions, enhancing processing speed by performing operations in parallel or sequentially, utilizing scene segmentation algorithms to determine the number of candidate regions based on the ROI size.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the system processes all regions of a large input image to ensure comprehensive object detection, then the detection coverage is improved, but the inference time increases significantly

Engineering Contradiction:
Improvedetection coverageVSAvoidinference time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent divides the input image into multiple regions of interest (ROIs) based on scene segmentation, focusing computational resources on areas containing objects of interest rather than processing the entire image uniformly. This segmentation approach maintains detection coverage while reducing the total number of regions requiring detailed processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system applies different processing strategies to different regions of the image based on their importance. Regions containing objects of interest receive detailed processing with multiple candidate regions generated, while less important regions receive simplified processing. This local quality differentiation improves inference speed without compromising detection reliability in critical areas.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If the system generates a high number of object candidate regions to improve detection accuracy, then the recognition precision is improved, but the processing complexity increases

Engineering Contradiction:
Improveobject recognition precisionVSAvoidprocessing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs preliminary scene segmentation to identify regions containing objects of interest before generating object candidate regions. By pre-identifying relevant areas, the system reduces the total number of candidate regions that need to be processed while ensuring that all actual objects are captured. This preliminary action maintains detection precision while reducing processing complexity.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If the system processes large input images with high resolution to improve object detection accuracy, then the detection precision is improved, but the computational load increases

Engineering Contradiction:
Improveobject detection accuracyVSAvoidcomputational load
Core Design Contradiction:
Measurement precisionVSPower

Solution Approach 1:

The patent extracts and processes only the relevant regions of interest from the large input image, rather than processing the entire high-resolution image. By extracting ROIs containing objects of interest and processing these smaller regions at high resolution, the system maintains detection accuracy while significantly reducing the computational load associated with processing the full image.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3451230B1Method and apparatus for recognizing object
Publication Date: 2024.09.18 SAMSUNG ELECTRONICS CO LTD
  • EP3451230B1 patent drawingFigure 1
  • EP3451230B1 patent drawingFigure 2
  • EP3451230B1 patent drawingFigure 3

AI summary

Methods and apparatus for recognizing an object are provided, including extracting a feature from an input image and generating a feature map in a neural network. In parallel with the generating of the feature map, a region of interest (ROI) corresponding to an object of interest is extracted from the input image, and a number of object candidate regions used to detect the object of interest is determined based on a size of the ROI. The object of interest is recognized from the ROI based on the number of object candidate regions in the neural network.