Explainable AI Verification for Autonomous Driving Image Reliability

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current autonomous driving systems rely on Deep Convolutional Neural Networks (CNNs) for image processing, but the reliability of input images is critical for safety, and existing technologies lack effective methods to verify image quality in real-time, posing a risk to safe autonomous driving.

Innovation Solution

A learning method and device using explainable AI to verify detection processes of neural networks for autonomous driving, which includes a property extraction module, a neural network for verification, and a loss module to assess image quality and predict safety information, generating notifications for the driver about potential dangers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If Deep CNNs are used for autonomous driving image processing, then object detection accuracy is improved, but reliability of image processing under abnormal conditions deteriorates

Engineering Contradiction:
Improveobject detection accuracyVSAvoidimage processing reliability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent introduces an intermediate verification module between the Deep CNN and the final output. This module includes a property extraction module that extracts image properties, a verification neural network that verifies detection results, and a contribution prediction module that predicts contribution values. These intermediaries act as mediators to check and validate the outputs of the Deep CNN, ensuring reliability even when input images are abnormal or detection results are incorrect.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If real-time image reliability assessment is implemented, then functional safety is improved, but system complexity increases

Engineering Contradiction:
Improvefunctional safetyVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the autonomous driving system into distinct functional modules: a Deep CNN for object detection, a property extraction module for image property analysis, a verification neural network for result verification, and a contribution prediction module for identifying problematic regions. Each module has a specific function and can be independently trained and optimized. This segmentation reduces system complexity by making each component manageable and modular while collectively achieving real-time reliability assessment.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The verification neural network and contribution prediction module use the output of the Deep CNN itself as input for verification. The system leverages its own detection results to self-validate, extracting properties from both the original image and the detection output. This self-service mechanism allows the system to assess its own reliability without requiring external verification systems, thereby reducing overall system complexity.

Inventive Principle:
Principle #25Self-service

3Reliability

If verification of detection processes is added, then safety warning capability is improved, but processing time increases

Engineering Contradiction:
Improvesafety warning capabilityVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The property extraction module performs preliminary analysis on the detection results by extracting key properties such as detection confidence, object attributes, and spatial information. This preliminary action prepares the data in a structured format that the verification neural network can quickly process. By pre-extracting relevant properties, the system avoids redundant computations during the verification stage, thereby reducing overall processing time while maintaining safety warning capability.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3690745B1Learning method and learning device for providing functional safety by warning driver about potential dangerous situation by using explainable ai which verifies detection processes of autonomous driving network, and testing method and testing device using the same
Publication Date: 2024.03.13 STRADVISION
  • EP3690745B1 patent drawingFigure 1
  • EP3690745B1 patent drawingFigure 2
  • EP3690745B1 patent drawingFigure 3

AI summary

A learning method for providing a functional safety by warning a driver about a potential dangerous situation by using an explainable AI which verifies detection processes of a neural network for an autonomous driving is provided. And the learning method includes steps of: (a) a learning device for verification, if at least one training image for verification is acquired, instructing a property extraction module to apply extraction operation to the training image for verification to extract property information on characteristics of the training image for verification to thereby generate a quality vector; (b) the learning device for verification instructing the neural network for verification to apply first neural network operations to the quality vector, to thereby generate predicted safety information; and (c) the learning device for verification instructing a loss module to generate a loss, and perform a backpropagation by using the loss, to thereby learn parameters included in the neural network for verification.